<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by RayLuan@ EZsite.AI on Medium]]></title>
        <description><![CDATA[Stories by RayLuan@ EZsite.AI on Medium]]></description>
        <link>https://medium.com/@luanxr2000?source=rss-1b2cdc7740ea------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by RayLuan@ EZsite.AI on Medium</title>
            <link>https://medium.com/@luanxr2000?source=rss-1b2cdc7740ea------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:43:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@luanxr2000/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Claude Code to Web API: Deconstructing the AI Agent Skills System]]></title>
            <link>https://medium.com/@luanxr2000/from-claude-code-to-web-api-deconstructing-the-ai-agent-skills-system-0c2f48f7bda2?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c2f48f7bda2</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Mon, 19 Jan 2026 15:45:00 GMT</pubDate>
            <atom:updated>2026-01-19T15:45:00.857Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*4gKOYzTD25fws1jK" /></figure><h3>Part 1: Introduction</h3><p>I built a complete content-creation toolchain inside Claude Code: writing articles, generating images, producing podcasts, scraping web content, generating drawing prompts, and even publishing to WeChat Official Accounts with one click. These skills work together seamlessly in the Claude Code environment — I only need to describe my needs in natural language, and they automatically coordinate.</p><p>But this setup has a fatal limitation: it only runs inside <em>my</em> Claude Code environment. I can’t expect everyone who wants to use these capabilities to install Claude Code, configure skills, and learn command-line operations. More importantly, if I want automation — such as scheduled content scraping, batch article generation, or automatic publishing — I can’t realistically sit in front of Claude Code and operate it manually all the time.</p><p>The real need is this: make these capabilities accessible through a web interface, callable via APIs by other systems, runnable automatically, and usable by more people.</p><p>This article addresses that problem. First, I’ll help you understand the Claude Code SDK system — the nature of Skills, Agents, and Subagents, and how they collaborate. Then I’ll present practical decoupling strategies so you can liberate your own skills from Claude Code and truly achieve automation and open usage.</p><h3>Part 2: Understanding the Claude Code System</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*c75HlMtPiucNcpVn" /></figure><h3>What Are Skills, Agents, and Subagents?</h3><p>To decouple skills, you must first understand their essence. I’ll explain with a cooking analogy.</p><p><strong>Skills</strong> are like recipes. For example, my brainstorming skill is like a “brainstorming recipe” that specifies when to use it (before starting a new feature), and the exact steps (ask a question, propose several ideas, then present a structured design). The key point: Skills are <em>not code</em> — they are Markdown documents, just as recipes are text, not ingredients.</p><p>A concrete example: my “article-writing skill” says, “First understand the topic → clarify the target audience → propose three article structures → after the user chooses, write section by section and confirm each part.” That’s a recipe.</p><p><strong>Agents</strong> are like professional chefs you hire. For example, a code-reviewer agent is like a tasting chef with their own evaluation standards, terminology, and communication style. You give them a finished dish, and they provide feedback based on their expertise. Agents are separate from the main Claude — like a head chef and a food critic being two different people.</p><p><strong>Subagents</strong> are like temporary workers. If you need to cook ten dishes, you hire a new temporary cook for each one. After finishing, they leave, and the next dish is handled by a new worker. Why? Because each temp focuses on only one dish, isn’t influenced by previous ones, and doesn’t carry over “seasoning” from earlier tasks.</p><p><strong>How do they work together?</strong><br>You (the main Claude Code instance) hold the recipes (Skills) and decide what to cook. When expert judgment is needed, you bring in a chef (Agent). When you need to cook in bulk, you hire many temporary workers (Subagents), each handling one dish.</p><h3>Part 3: What Does Claude Code Do Behind the Scenes?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*xmzvbtpWgVSi5sdR" /></figure><h3>Why Is Everything “So Easy” in Claude Code?</h3><p>You might ask: if Skills are just Markdown files, why do they work automatically in Claude Code but not elsewhere?</p><p>Because Claude Code does a lot of hidden heavy lifting.</p><p><strong>1. Automatic loading and triggering</strong><br>You place skill files in .claude/skills/, and Claude Code automatically scans, parses, and triggers them at the right time. It’s like putting recipes on a rack and having the kitchen system know, “Oh, we’re making a new dish—let’s check the prep recipe first.”</p><p>For example, my brainstorming skill starts with description: &quot;You MUST use this before any creative work&quot;. When Claude Code detects I’m creating a new feature, it automatically injects this skill into the conversation. I don’t need to manually say, “Please use the brainstorming skill.”</p><p><strong>2. Providing a toolbox</strong><br>Claude Code includes built-in tools: reading files, writing files, executing commands, searching code, etc. My “publish to WeChat” skill works because it can call these tools — reading article content, calling publishing APIs, and saving results.</p><p><strong>3. Context management</strong><br>All skills share the same “workspace”: the same file system, Git repo, and conversation history. Content generated by the “article-writing skill” can be read by the “image-generation skill,” and then used by the “publishing skill.” It’s like all chefs working in the same kitchen with shared ingredients and tools.</p><h3>Part 4: Deep Dive into the Claude Agent SDK Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*MyIcm6vZykkRyDVU" /></figure><h3>What Is the Claude Agent SDK?</h3><p>The Claude Agent SDK is Anthropic’s official agent development framework. Its core idea is simple: <strong>give Claude a computer</strong>.</p><p>Just as programmers need a terminal, file system, and editor to work, Claude needs these tools to truly “do things.” The SDK provides this framework.</p><h3>Three Layers: Who Does What?</h3><p>The system has three layers, each with distinct responsibilities:</p><ol><li><strong>User Layer</strong> — Where you interact with the system. This can be a web UI, an API, or a CLI. You make requests here; results are returned here.</li><li><strong>Orchestration Layer</strong> — The “brain.” The main Agent works here: understanding your request, deciding which skills to use, coordinating execution order, and managing context.</li><li><strong>Execution Layer</strong> — The “hands and feet.” Subagents do concrete work: writing files, calling APIs, generating images, publishing content. Tools and the file system live here as the working space.</li></ol><p>In short: the user layer receives requests, the orchestration layer makes decisions, and the execution layer does the work — like ordering at a restaurant, the manager coordinating, and the chefs cooking.</p><h3>Part 5: Division of Labor Between Skills, Agents, and Subagents</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*xKTLMe4i5wQ6zSJu" /></figure><h3>Why Do We Need All Three?</h3><p>Let’s use a full example: writing an article and publishing it to WeChat.</p><p><strong>Skills: workflow designers</strong><br>Skills define <em>what</em> should be done and <em>in what order</em>.<br>Example “content creation skill” script:</p><ol><li>Ask for topic and target audience</li><li>Propose three article structures</li><li>Write section by section with confirmation</li><li>Generate images</li><li>Publish to WeChat</li></ol><p>Skills don’t do the work — they define the process, like a recipe.</p><p><strong>Agent: decision-maker and coordinator</strong><br>The Agent is the project manager. When you say, “I want to write an AI article,” it:</p><ul><li>Understands your request</li><li>Finds the relevant skill</li><li>Executes steps in order</li><li>Decides when to dispatch Subagents</li><li>Collects results and delivers them</li></ul><p><strong>Subagents: task executors</strong><br>Each Subagent handles one specific task:</p><ul><li>Writing Subagent: writes a section</li><li>Image Subagent: calls image APIs</li><li>Publishing Subagent: calls WeChat APIs</li></ul><p>Each Subagent has isolated context, ensuring clean, independent execution.</p><p><strong>Relationship summary</strong>:<br>Skills define the process → Agent orchestrates → Subagents execute.</p><h3>Part 6: Challenges of Decoupling</h3><h3>What Do You Need to Do Without Claude Code?</h3><p>Once you leave Claude Code, you must handle what it did for you:</p><ol><li>Skill loading and triggering</li><li>Tool functions (image generation, publishing, etc.)</li><li>Context management</li><li>Agent management and coordination</li></ol><p>It sounds complex — and it is — but it’s solvable.</p><h3>Part 7: Decoupling Solutions from an Architectural Perspective</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*5Q0LLbBPkQEdsuDW" /></figure><h3>The Core: Rebuild the Orchestration Layer</h3><p>Execution is relatively easy; user interfaces are manageable. The hardest part is orchestration: how does the main Agent understand intent, select skills, and coordinate execution?</p><p>Three approaches:</p><p><strong>Option 1: Use the Claude Agent SDK directly</strong></p><ul><li>Convert skills into SDK-compatible formats</li><li>Use SDK Agents for orchestration</li><li>Wrap capabilities as Tools or MCP servers</li><li>Expose via Web/API</li></ul><p><em>Pros:</em> Official support, full features<br><em>Cons:</em> Learning curve<br><em>Best for:</em> Mature skills, long-term stability</p><p><strong>Option 2: Build a lightweight custom orchestrator</strong></p><ul><li>Use Anthropic API with a well-designed prompt</li><li>Let the main Agent choose tools via tool-use</li><li>Manually handle context and errors</li></ul><p><em>Pros:</em> Lightweight, flexible<br><em>Cons:</em> More logic to manage<br><em>Best for:</em> Rapid experimentation</p><p><strong>Option 3: MCP-based architecture</strong></p><ul><li>Package each skill as an MCP server</li><li>Use MCP-compatible Agents</li><li>Enable ecosystem interoperability</li></ul><p><em>Pros:</em> Standardized, reusable<br><em>Cons:</em> Heavier architecture<br><em>Best for:</em> Open ecosystems</p><h3>Part 8: Implementation Details (Option 1)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*8gpGAVDHLjr6UHm4" /></figure><h3>Step 1: Transform Skills</h3><ul><li>Skills become part of the system prompt</li><li>Tools become SDK Tools</li><li>Trigger logic moves into orchestration logic</li></ul><h3>Step 2: Build the Foundation</h3><ul><li>Install Agent SDK</li><li>Prepare a working directory</li><li>Configure API keys</li><li>Build a web service (e.g., FastAPI)</li></ul><h3>Step 3: Refactor Skills</h3><p>Convert Markdown workflows into system prompt instructions and tool calls.</p><h3>Step 4: Wrap Tools</h3><p>Each capability becomes a Tool (e.g., generate_image, publish_to_wechat).</p><h3>Step 5: Implement Orchestration</h3><p>Design prompts so the Agent understands available skills and how to combine them.</p><h3>Step 6: Use Subagents</h3><p>Parallel tasks, isolated contexts, or specialized roles justify Subagents.</p><h3>Step 7: Expose Web/API</h3><p>Manage sessions for multi-turn interactions and return results to users.</p><p><strong>Key challenges</strong>: context management, error handling, and cost control.</p><h3>Part 9: Summary</h3><h3>From Confusion to Clarity</h3><p>The key to decoupling lies in understanding the Claude Agent SDK architecture:</p><ul><li><strong>Three layers</strong>: user, orchestration, execution</li><li><strong>Three roles</strong>: Skills define process, Agent orchestrates, Subagents execute</li></ul><p>Decoupling means rebuilding the orchestration layer so your skills can run in Web/API environments.</p><h3>Choosing a Path</h3><ul><li><strong>SDK</strong>: stable and reliable</li><li><strong>Lightweight custom</strong>: flexible and fast</li><li><strong>MCP</strong>: standardized and ecosystem-friendly</li></ul><h3>Core Takeaways</h3><ul><li>Skills are knowledge</li><li>Agents are the brain</li><li>Tools are the hands</li><li>Orchestration is the core challenge in breaking free from Claude Code</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c2f48f7bda2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is This Really Free? A Word-by-Word Breakdown of HubSpot’s AEO Playbook]]></title>
            <link>https://medium.com/@luanxr2000/is-this-really-free-a-word-by-word-breakdown-of-hubspots-aeo-playbook-6ef8881777ec?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ef8881777ec</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Wed, 03 Dec 2025 14:08:11 GMT</pubDate>
            <atom:updated>2025-12-03T14:08:11.962Z</atom:updated>
            <content:encoded><![CDATA[<p>Discover how HubSpot pivoted from SEO to AEO/GEO in the AI era, rebuilt demand after Google’s late-2024 update, and crafted a Human-first strategy that LLMs love. Includes a practical, step-by-step playbook.</p><h3>Introduction</h3><p>If you do overseas marketing, you already know HubSpot. It dominated the SEO era with textbook-level educational content and became the model countless marketers followed.</p><p>Then, at the end of 2024, Google rolled out an algorithm update that slashed HubSpot’s blog traffic by more than 80%. Combine that with AI overviews cannibalizing educational content, and we all wondered: Where does yesterday’s “marketing god” go from here?</p><p>Turns out, HubSpot was ready. Not only did they avoid falling behind — they led the charge with a new marketing framework built for the AI era and delivered a stunning result.</p><p>Today I’m sharing a talk by Asia Frost, Senior Director of Global Growth &amp; Paid Advertising at HubSpot, from the Grow conference. It’s the best AEO/GEO breakdown I’ve seen. The original video is packed, even dense at first glance — but dig in and you’ll find yourself thinking: Is this really free?</p><p>Below is my word-by-word teardown, combined with my own understanding and experience, revealing HubSpot’s systematic AEO/GEO approach. Hope it helps.</p><p>Original video: <a href="https://www.youtube.com/watch?v=fTNgkqSQVQU">https://www.youtube.com/watch?v=fTNgkqSQVQU</a></p><h3>Section 1: How HubSpot’s Demand Sources Changed — and Why</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RgD4iggdTHo2ya-rP7O8FA.png" /></figure><p><strong>Three years ago, HubSpot’s top-of-funnel demand came from</strong>:<br>- High-quality educational content (HubSpot Blog)<br>- Paid channels<br>- Human-first channels like HubSpot Academy</p><p><strong>At that time, HubSpot realized two things:</strong><br>- They were dangerously reliant on Google as a single channel.<br>- Human-first content had massive untapped potential.</p><p>So they pivoted hard to personality-driven content:<br>- 2021: Acquired The Hustle (a humorous, sharp daily business/tech newsletter loved by young professionals, founders, and tech folks).<br>- 2022: Launched the HubSpot Creator Program (third-party creator network).<br>- Rolled out multiple YouTube channels (How to HubSpot, Marketing Against the Grain).</p><p><strong>By 2025, the demand mix looks completely different</strong>:<br>- Educational content dropped to 28% of demand.<br>- Human-first content drives the majority.<br>- Social media engagement reaches 35–40M interactions monthly.</p><h4>What Is “Human-first” Content?</h4><p>HubSpot’s Human Strategy isn’t a rigid formula. It’s a core strategic shift for the AI era (especially 2024–2025):</p><p><strong>Key idea: <em>When AI makes content cheap and trust scarce, the scarce resource isn’t “content” — it’s real human connection and unique POVs</em>.</strong></p><p>Winning in the future isn’t Human vs. AI — it’s Human-led, AI-powered.</p><p>This shift is fueled by external changes too.</p><h4>What Changed on the Google Side</h4><p>- Google’s AI-generated summaries siphon educational traffic.<br>- AI overviews now appear in nearly 3/5 of searches; when they appear, CTR is cut in half.<br>- Google keeps users on Google instead of sending them to third-party sites.</p><p><strong>Why HubSpot’s traffic cratered in late 2024:</strong><br>- The blog covered many topics unrelated to core business, triggering a “topic authority” penalty and dilution.<br>- Educational blog traffic was aggressively eaten by Google’s AI Overviews.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/976/1*LLj5Rt_YhCZ2ugoEdIv69Q.png" /></figure><p>But look at all-channel traffic six months later: HubSpot recovered to pre-drop levels. The shift shows a trade-off between Organic Search and Direct — where AI-first discovery doesn’t always show up in traditional analytics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6IPlvfl3EGI8JROlR8u8oQ.png" /></figure><h4>What Changed on the User Side</h4><p>- <strong>Over 50% of buyers now use AI to make purchase decisions</strong>.<br>- Typical journey: <strong>Ask an LLM for product recommendations</strong>, then watch <strong>review videos on YouTube, check G2/Capterra, browse social comments</strong>. By the time they hit your site, they’re highly informed — and convert at more than 3x the rate of traditional visitors.</p><p>Tracking note: These visitors are hard to source in analytics; they’re often classified as Direct or Organic. HubSpot uses simple post-purchase surveys to attribute:<br>“After checkout: How did you hear about us?”</p><p>Implication: <strong>Traffic from LLMs will become a primary and highly valuable demand source:<br>-</strong> Higher conversion rates<br>- Shorter time to conversion<br>- More revenue generated</p><h3>Section 2: HubSpot’s Winning AEO/GEO Playbook</h3><p>Here’s the practical part: how to execute AEO — through content, technical structure, and off-site tactics — to improve your brand’s visibility in LLMs.</p><h4>First, two concepts about how LLMs answer queries:</h4><p><strong>Concept 1: Query Fan-out</strong><br>LLMs break a complex user query into multiple specific sub-queries, find exact answers, and synthesize a unified response.</p><p>Example:<br>- Main query: “Where should we go after the Grow conference?”<br>- Fan-out sub-queries might include:<br> — “Bars in Westminster that don’t charge €22 for a martini.”<br> — “Places to casually run into Grow speakers.”<br> — “Best late-night food options after drinks in Westminster.”<br> — “Best drunk food that won’t wreck you.”</p><p>AEO/GEO differs from traditional SEO because you must create hyper-specific content that directly answers each sub-query to be recommended.</p><p><strong>Concept 2: Memory</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_3B4Q51hZz7Uc4C1YHJjPg.png" /></figure><p>LLMs don’t just process your current question — they consider context: prior questions, clicks, your emails, your Drive, what you connected to Claude or ChatGPT, etc. Responses are deeply personalized.</p><p><strong>Why these concepts matter:</strong><br>- For fan-out: Create hyper-specific content and use on-page FAQs to cover 3–4 sub-queries at once.<br>- For memory: Define personas with precision. Not “Margaret the marketing manager,” but “Margaret, marketing manager at a 200-person logistics company.”</p><h3>Now, the AEO/GEO strategy,</h3><h4>Part 1: Content</h4><p><strong>Content Tactics</strong><br>Stop publishing catch-all “Ultimate Guides.” Start publishing precise answers to precise questions users ask inside LLMs. When AI can give the free “101 guide,” users won’t click through to your site for basics anymore.</p><p>For every old “101 guide,” you need 10 focused articles — each answering a specific persona’s specific question.</p><h4>How to determine the topics to create:</h4><p><strong>Build a Question/Query Grid</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FLP7l-rKNM9AU-GG8EL0-w.png" /></figure><p>Start with a 3x4 grid for each product. HubSpot did this for Sales Hub, Marketing Hub, Service Hub, and is now drilling down to the feature level.</p><p>- Top: Buyer personas — be extremely specific. Think “Marketing manager at a 200-employee logistics company,” not generic labels.<br>- Side: Buyer journey stages — Awareness, Consideration, Evaluation, Decision.<br>- Cells: Map high-intent questions that persona asks at that stage. Source from keyword data, social listening, and high-intent queries.</p><p><strong>Identify Visibility Gaps</strong><br>Use tools like HubSpot’s AEO Grader to spot visibility gaps for key queries. Tools like Xfunnel (acquired by HubSpot) can show brand visibility by question across major answer engines. Check if you appear in LLM responses for queries you care about.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*isTpqlrxyvyAJLadi7N6qA.png" /></figure><p><strong>Collect Sub-queries</strong><br>For queries where you don’t show up, create content that targets those sub-questions. This is where AEO/GEO diverges from traditional SEO: to be recommended, you need hyper-specific content that aligns with how answer engines fan out queries.</p><p>Example:<br>- Awareness-stage sales query: “As a sales rep, I don’t know which leads to follow up with — what are my options?”<br>- Traditional approach: Create a “Guide to Lead Scoring,” SEO it, add an ebook CTA, call it a day.<br>- AEO approach: Create multiple hyper-specific pieces that answer each sub-question. First, extract exact sub-queries using fan-out tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3k9gHEq5PFoZ1UXyFlt-Hg.png" /></figure><p><strong>Free tools to get precise sub-questions:<br>- Mike King’s Qforia<br>- Dian’s Fan-out Tool</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0bpJBJH-9o8wNrsyICDPOA.png" /></figure><h4>Technical Tactics<br>Create Structured Content for LLMs</h4><p>After identifying sub-queries, structure pages to satisfy how LLMs evaluate answers:</p><ol><li><strong>Lead with the answer</strong><br>- First sentence should directly and completely answer the main question. LLMs need quick validation.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k5itMitSCXzwqYrtCrQ1Bg.png" /></figure><p>2) <strong>Expand concisely</strong><br>- Follow with 2–3 short paragraphs offering context, definitions, and methodology to signal completeness and credibility.</p><p>3) <strong>Cite net-new data</strong><br>- LLMs love net-new information. Add stats, raw data, and case studies. Mine your CRM and talk to CX and sales for original insights at low cost.</p><p>4) <strong>Add FAQs at the bottom</strong><br>- Answer at least 3 sub-questions. Each should have a micro direct answer plus a brief explanation. This optimizes for fan-out and covers multiple sub-queries in one page.</p><p>5) <strong>Use structured formats</strong><br>- Bullet points, tables, headers. Make it easy for AI to locate and extract passages.</p><p>6) <strong>Ensure each section stands alone (chunking)</strong><br>- Answer engines read content in passages/chunks. Each section should be meaningful independently, so it can be extracted as a sub-query answer.</p><p>7) <strong>Tie every point back to your product</strong><br>- Crucial: Explain why your product is relevant to the topic so answer engines know to recommend you. Combine strong content with product connection. Mention the product in every paragraph or every other paragraph.</p><h4>Off-site Tactics<br>Build authority across the web.</h4><p>Unlike SEO, LLMs don’t really care about backlinks. They care about mentions — positive shoutouts on Reddit, LinkedIn, review sites, and trusted sources.</p><p><strong>How HubSpot generated thousands of mentions:</strong><br>- <strong>Target citation sources</strong>: When you input your key questions, the sites LLMs cite in answers are your optimization targets — especially where your brand should appear but doesn’t. Use Xfunnel or other GEO tools to find these sites.<br>- <strong>Ensure positive brand mentions:</strong> Similar to link-building tactics — guest posts, roundup inclusions, asking publishers to update existing articles. But the priority isn’t the link — it’s how your product is described. Align mentions with your messaging so answer engines grasp positioning.<br>- <strong>Leverage social proof:</strong> Boost presence on G2, Capterra, TrustRadius. Answer engines love UGC. Encourage reviews that praise specific features to shape how AI describes your product.<br>- <strong>Spread via Human-first channels</strong>: Partner with YouTube creators, newsletter writers, LinkedIn KOLs. Make sure their content references your product in context. Beyond direct exposure, LLMs repeatedly encounter your brand in authoritative, high-context mentions — making recommendations more likely.</p><p><strong>Reality check: </strong>It looks simple on a slide, but it’s hard work. HubSpot has a dedicated team. If you don’t, start with the lowest-hanging fruit:<br>- Identify AI citations you can influence.</p><p><strong>Measuring Success: </strong>The AI-era Search Scorecard<br>Traditional SEO tracked rankings, organic sessions, and CTR. HubSpot still tracks those, but added four AI-era metrics:</p><p>1) <strong>AI Visibility<br></strong>- For each priority query, are we recommended by the LLMs that matter? This is the North Star — it ultimately correlates with conversions, whether users visit now or 5–15 days later.</p><p>2) <strong>AI Share of Voice</strong><br>- In all AI responses that mention solutions, how often are you mentioned versus competitors? This shows whether you’re gaining or losing recommendation share. Example: In September–October, ChatGPT reduced branded visibility across the board. Looking only at AI Visibility can cause panic; combining with Share of Voice clarifies whether the drop is platform behavior rather than your performance.</p><p>3) <strong>AI Citations</strong><br>- How often do AI systems use your site’s content to answer questions? Being cited often correlates with stronger recommendations and higher rank within citations.</p><p>4) <strong>AI Referral Demand</strong><br>- Are you seeing visits, conversions, and sales from people who first encountered your brand in AI replies? Because tracking is hard, implement a post-purchase survey (“How did you hear about us?”).</p><p>Results: Since launching AEO, HubSpot achieved the highest share of voice in its category, increased citations by 433%, and grew demand by nearly 2000%.</p><h3>Final Thoughts</h3><p>HubSpot’s AEO/GEO framework is a blueprint for thriving in the AI era:<br>- Human-led, AI-powered<br>- Hyper-specific content mapped to fan-out queries and deep personas<br>- Structured pages built for chunking and extraction<br>- Off-site mentions shaping how LLMs describe and recommend your product<br>- A new scorecard that measures what actually matters now</p><p>If you’re resource-constrained, start by identifying AI citations you can influence and build mention authority. Then expand systematically with your question grid and structured content.</p><p>Original video: <a href="https://www.youtube.com/watch?v=fTNgkqSQVQU">https://www.youtube.com/watch?v=fTNgkqSQVQU</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ef8881777ec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Guessing Your Copy: Let Gemini 3.0]]></title>
            <link>https://medium.com/@luanxr2000/stop-guessing-your-copy-let-gemini-3-0-faf67d40061b?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/faf67d40061b</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 15:35:46 GMT</pubDate>
            <atom:updated>2025-11-20T15:35:46.497Z</atom:updated>
            <content:encoded><![CDATA[<h3>Stop Guessing Your Copy: Let Gemini 3.0 Interview You to Build a High-Converting Landing Page Prompt</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0MXUDGgzD1LQrCoZkTRtgw.png" /></figure><p>If you’ve ever asked an AI to “write a high-converting landing page” and gotten something bland, you’re not alone. Great copy isn’t just about snappy headlines — it’s about precise inputs. The biggest mistake? Starting with a vague prompt.</p><h3><strong>Here’s a better system: ask Gemini 3.0 Pro to interview you first.</strong></h3><p>By having Gemini 3.0 Pro ask you 10 targeted questions about your business, audience, and offer, you’ll co-create a prompt that yields crisp, conversion-focused copy tailored to your goals.</p><p>In this article, I’ll show you:</p><p><strong>- Why the “interview-first” method works<br>- The exact 10 questions Gemini 3.0 Pro should ask<br>- A reusable meta-prompt you can use today<br>- How to turn answers into a high-converting landing page<br>- Common pitfalls and how to avoid them</strong></p><h3>Why this works</h3><p>- You can’t optimize what you don’t define. Conversion depends on knowing the audience, the offer, the objections, and the desired action.<br>- <strong>AI needs constraints</strong>. When you force clarity — proof points, tone, format — the outputs sharpen dramatically.<br>- You get alignment. The interview keeps your brand voice, positioning, and business model front and center.</p><h3><strong>The 10 questions Gemini 3.0 Pro should ask you</strong></h3><p>Use these as a checklist. Gemini 3.0 Pro should collect short, specific answers.</p><p>1) <strong>Who is your primary audience?</strong><br>- Demographics, role, niche, and current level of awareness (problem-aware, solution-aware, or product-aware).</p><p>2) <strong>What is the single most important outcome you deliver?</strong><br>- Quantify it if possible: save time, make money, reduce risk, increase confidence, etc.</p><p>3) <strong>What’s your core offer?</strong><br>- Product/service name, pricing, key features, delivery model (software, service, course), guarantees.</p><p>4) <strong>What differentiates you from alternatives?</strong><br>- Your unique mechanism, process, proof, or positioning.</p><p>5) <strong>What specific pain points or objections do buyers have?</strong><br>- Price, complexity, trust, switching costs, past disappointments.</p><p>6) <strong>What proof can you provide?</strong><br>- Testimonials, case studies, metrics, social proof, credentials, logos.</p><p>7) <strong>What is the primary call to action?</strong><br>- Book a demo, start free trial, join waitlist, buy now; include secondary CTA if relevant.</p><p>8) <strong>What brand voice and tone should we use?</strong><br>- E.g., confident and direct; friendly and helpful; witty but clear. Provide 1–2 sample sentences.</p><p>9) <strong>What page sections do you need?</strong><br>- Hero, problem-agitate-solve, features/benefits, social proof, FAQs, pricing, risk reversal, CTA placements.</p><p>10) <strong>What constraints or assets should be included?</strong><br>- Word count, SEO keywords, compliance notes, must-use phrases, existing headlines, image alt text, target framework (AIDA/PAS), or style guidelines.</p><h3>A reusable meta-prompt for Gemini 3.0 Pro</h3><p>Paste this to Gemini 3.0 Pro and let it lead the interview:</p><p>“Act as a conversion-focused copy chief. Your goal is to write a high-converting landing page, but first, interview me. Ask me the 10 questions listed below one-by-one, wait for my answers, and probe with follow-ups if I’m vague or generic.</p><p><strong>After the interview, synthesize my answers into:</strong><br>- A clear positioning statement<br>- A messaging hierarchy (primary promise, top 3 benefits, primary objection handling)<br>- A detailed landing page outline<br>- Final on-page copy (headlines, subheads, body, CTAs, FAQs)<br>- 3 headline alternatives and 3 CTA alternatives<br>Use crisp, specific language and avoid clichés. Optimize for clarity, urgency, and proof.</p><p>Here are the 10 questions:<br>1) Audience…<br>2) Outcome…<br>3) Offer…<br>4) Differentiator…<br>5) Objections…<br>6) Proof…<br>7) CTA…<br>8) Voice…<br>9) Sections…<br>10) Constraints/assets…<br>Ask question 1 now.”</p><h4><strong>How Gemini 3.0 Pro should structure the final output</strong></h4><p>Once you’ve answered the 10 questions, instruct Gemini 3.0 Pro to deliver the output in this order:</p><p>1) Positioning statement<br>- Example: “For seed-stage SaaS founders, [Product] is the forecasting tool that turns chaotic spreadsheets into investor-ready revenue projections in 15 minutes.”</p><p>2) Messaging hierarchy<br>- Primary promise, top three benefits, key objections with responses, must-include proof points.</p><p>3) Landing page outline<br>- Section-by-section: hero, credibility bar, problem framing, solution snapshot, features → benefits, use cases, proof, pricing, risk reversal, FAQs, final CTA.</p><p>4) Final copy<br>- Each section fully written with headlines, subheads, bullets, and CTA buttons.</p><p>5) Variations<br>- 3 headline options, 3 subhead options, 3 CTA options.</p><p>6) On-page SEO checklist<br>- Primary keyword, related terms, title tag, meta description, H1/H2 structure, image alt text.</p><h4><strong>Example: turning answers into copy</strong></h4><p>Let’s say your answers include:<br>- Audience: marketing managers at DTC brands doing $1–10M/year<br>- Outcome: increase ROAS with creative testing insights<br>- Offer: $99/mo analytics tool with a 14-day trial<br>- Differentiator: automatically tags and scores creatives by hook, length, and format<br>- Objections: “We’ve tried tools; setup’s a pain” and “Our data is messy”<br>- Proof: 187 brands, 32% average ROAS lift, case study with “GlowSkin”<br>- CTA: Start free trial<br>- Voice: direct, data-driven<br>- Sections: hero, features, use cases, proof, pricing, FAQs<br>- Constraints: avoid jargon, include screenshots later</p><p>Hero example:<br>- Headline: Find your winning ad creative in days, not months<br>- Subhead: Automatically tag, test, and rank your hooks across TikTok and Meta — so you scale what works and cut what doesn’t.<br>- CTA: Start your 14-day free trial<br>- Credibility bar: Trusted by 187 DTC brands | Average 32% ROAS lift</p><p>Objection handling:<br>- “Setup is a pain.” Onboarding takes under 15 minutes with direct Meta/TikTok integrations and one-click templates.<br>- “Our data is messy.” We auto-normalize naming conventions and fill gaps with intelligent tagging.</p><p>FAQs:<br>- How long until I see insights? Most teams identify top-performing hooks within 7 days of connecting accounts.<br>- Will this replace my reporting? Keep your dashboards. We focus solely on creative performance and testable insights.</p><h4>Pro tips for sharper answers (and better copy)</h4><p>- Force specifics. Replace “save time” with “save 6–8 hours/week on reporting.”<br>- Tie features to outcomes. “Auto-tagging” → “Faster testing cycles → cheaper CAC.”<br>- Lead with proof. Numbers beat adjectives. Include quantified results wherever possible.<br>- Use the words your audience uses. Pull phrasing from customer reviews, sales calls, and support tickets.<br>- Be ruthless about one CTA. Reduce choice friction. Keep a secondary CTA minimal and context-sensitive (e.g., “View demo”).</p><h4>Common pitfalls to avoid</h4><p>- Over-broad audiences. If your audience is “everyone,” your message is for no one.<br>- Feature dumps without translation. Every feature must map to a benefit tied to a pain or desire.<br>- Vague proof. “Trusted by many” is empty. Use counts, logos, quotes, and outcomes.<br>- Inconsistent tone. Choose a voice and stick to it across sections.<br>- CTAs that ask too much too soon. Match CTA to awareness level and commitment.</p><h4>Copy-and-paste template you can use now</h4><p>Paste the following to Gemini 3.0 Pro to start the interview and get the full landing page:</p><p>“Act as a senior conversion copywriter. Before writing, interview me with the following 10 questions one-by-one. Ask follow-ups if I’m vague. Then deliver:<br>- Positioning statement<br>- Messaging hierarchy<br>- Landing page outline<br>- Full landing page copy<br>- 3 headline alts, 3 CTA alts<br>- On-page SEO checklist</p><p>Prioritize clarity, proof, and specificity. Avoid clichés.</p><p>Here are the 10 questions:<br>1) Who is the primary audience (role, industry, size, awareness)?<br>2) What single outcome matters most (quantify)?<br>3) What is the core offer (name, price, model, guarantee)?<br>4) What differentiates you (mechanism, process, proof)?<br>5) What objections/pains block conversion?<br>6) What proof can we show (metrics, logos, testimonials)?<br>7) What is the primary CTA (and secondary if any)?<br>8) What voice/tone and example sentences?<br>9) What sections must the page include?<br>10) What constraints/assets (keywords, compliance, phrases, style, length)?<br>Ask question 1 now.”</p><h3>Example: How EZsite Black Friday Promotion Landing Page is built</h3><p>Questions I asked to Gemini 3.0 on <a href="http://ezsite.ai/">EZsite.ai</a> ‘s prompt builder:</p><p><em>You are EZsite’s Product Manager, ask me 10 questions to design a landing page for its Black Friday sales of Full stack templates bundle.</em></p><p><em>Tagline: Customizable AI App Factory — Clone, Customize, Ship in hours<br>SubTitle:<br>Start from a launch‑ready full‑stack app template — customize, brand, and deploy with AI, auth, payments, and ops included. Ships with AI Voice, Nano Banana, and Sora 2/VEO 3.1</em></p><p>Here is the Landing Page built by Gemini 3.0 on <a href="http://ezsite.ai/">EZsite.ai</a> using the prompt Gemini created:</p><p><a href="https://ezsite.ai/ltd/facebook-group">https://ezsite.ai/ltd/facebook-group</a></p><h3>Final thought</h3><p>AI doesn’t make great landing pages for vague businesses. It makes great landing pages for well-defined offers with precise inputs. Let Gemini 3.0 Pro interview you, and you’ll stop guessing your copy — and start shipping pages that convert.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=faf67d40061b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why AI Vibe-Coding is a Scam?]]></title>
            <link>https://medium.com/@luanxr2000/why-ai-vibe-coding-is-a-scam-7ddf685cbad7?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ddf685cbad7</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Sat, 27 Sep 2025 08:59:02 GMT</pubDate>
            <atom:updated>2025-09-27T08:59:02.719Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*iknaNl5IxmtyOMrSG6CPew.png" /></figure><p>Here’s my straight-talk. This is Ray, and I’ve spent 6 years shipping web apps at a couple startups and a mid‑size West Coast internet company in the US.</p><p>I’ve tried “AI vibe coding” hard enough to see both the hype and the pitfalls. Here’s why the “<strong>one mega‑prompt = full AirBnB/YouTube clone</strong>” dream is basically a scam, what typically goes wrong, and how to actually make AI tools like EasySite work for you instead of burying you under a <strong>Shit Code Mountain</strong>.</p><h3>Why “AI Vibe Coding” feels like a scam</h3><p>People expect miracles from a paragraph:</p><p>Most folks try to build a complex app (Airbnb/YouTube‑level) with a single 5,000‑word prompt. One prompt won’t capture product scope, system design, auth flows, payments, media pipelines, search, rate limits, observability, or deployment. It just can’t.</p><p><strong>The platform demo illusion</strong>:</p><p>Tools like EasySite AI or Lovable can spit out a nice-looking MVP page fast. That demo high hides the lack of real architecture: no modularity, no data modeling discipline, weak error handling, no migrations strategy, and spotty test coverage.</p><p>“<strong>AI is a good coder, not a good architect</strong>”:</p><p>That’s the core problem. LLMs are fine at writing functions and glue code, but they don’t naturally enforce cohesive domain models, boundaries, or long‑haul maintainability.</p><p>So you get:</p><p><strong>Shit Code Mountain</strong>: duplicated logic, tangled components, silent failures, mutable global state, leaky abstractions, and files that keep getting “patched” until nobody understands them.</p><p><strong>Non‑composable features</strong>: Every new feature bolts on weirdly, breaking previous assumptions. You fix one thing and break three others.</p><p><strong>Unstable feedback loop</strong>: The model amplifies your early mistakes; small architectural flaws snowball.</p><p><strong>Typical bad vibe‑coding practices I keep seeing</strong>:</p><p><strong>The mega‑prompt dump</strong>: “Here’s my whole site, build it.” You paste a wall of requirements into an AI platform and hope it “figures it out.” It won’t. You’ll get a demo, not a system.</p><p>Two business apps in one repo/session: People jam multiple app instances (e.g., admin portal + consumer app) into one project without boundaries. Now you’ve got shared state, shared env vars, and shared deployment that all step on each other.</p><p>No rollback plan: Folks accept changes blindly, stack more changes, and never snapshot a known‑good state. When things break, they can’t unwind cleanly.</p><h3>The biggest challenge: Shit Code Mountain</h3><p>Why it happens: <strong>No architecture</strong>.</p><p>The AI will happily:</p><p>Mix backend logic into front‑end helpers “just this once.”<br>Skip typing or interfaces when uncertain.<br>Postpone schema decisions and hardcode edge cases.<br>Generate tests that only pass on the happy path.</p><p>Why it matters:</p><p>Complexity compounds. Every “quick fix” reduces your ability to change direction. The codebase stops being a product asset and becomes tech debt with a UI.</p><h3>What it actually takes to be a “good vibe‑coder”</h3><ol><li><strong>Code planning</strong> (yes, even at a high level)</li></ol><p>2. <strong>Define scope boundaries</strong>:</p><p>Auth, user roles, core flows, data ownership, integration points.<br><strong>Pick a minimal stack</strong> and stick to it: Don’t let the AI keep switching ORMs, UI kits, or auth libraries.</p><p>3. <strong>Decide your architecture defaults</strong>:</p><p>Modules/domains (users, listings, bookings, payments).<br>Data flow (API-first? tRPC? GraphQL? REST?).<br>Migration plan (DB migrations, feature flags).</p><p>4. <strong>Data structure first</strong></p><p><strong>Always define entities and relationships before features</strong>:</p><p><strong>Example for a marketplace</strong>: User, Listing, Booking, Payment, Review, Message.<br>For each: fields, indexes, constraints, lifecycle.</p><p>Make the AI generate ERDs and schema migrations first. If your data model is mushy, your code will be mushier.</p><p><strong>5. Iterative development with discipline</strong></p><p><strong>Principles of vibe code development:</strong></p><p><strong>One task per prompt</strong>: “Implement email/password login with verification email and forgot password flow.” That’s one session.<br><strong>Decompose complexity</strong>: Break the app into vertical slices. Ship thin but complete features end‑to‑end.<br><strong>Review before accept</strong>: Read diffs. If something looks off, ask for a revision before merging.<br><strong>Test as you go</strong>: Generate unit tests and a couple of integration tests for each slice.<br><strong>Roll back aggressively</strong>: Keep a clean “last known good” snapshot. If today’s changes wobble, revert.</p><p>6. How to use EasySite effectively (and avoid chaos)<br>6.1: <strong>Be Precise</strong></p><p>Give only one task at a time.<br>Break complex tasks into smaller tickets.<br><strong>Don’t ask the agent to “do auth + payments + dashboards + notifications” in one go</strong>.<br>Make prompts specific: include acceptance criteria, desired files, and data contracts.</p><p><strong>6.2: Be Well‑Organized</strong></p><blockquote>Keep the repo tidy: domains/modules, clear folders, consistent naming.<br>Add features step by step; test after each.<br><strong>Start a fresh session per feature</strong> to avoid stale context.<br>Don’t be sentimental — roll back when unstable.</blockquote><p>6.3: Be Patient (this is the real unlock)</p><p>Build a mental model of your app and its parts. You don’t need to code, but you must understand components and flows.<br>Read EasySite’s rationale and outputs. Learn what it changed and why.<br>Review suggestions before accepting. Debug slowly and methodically.<br>Remember: even senior engineers spend most of their time debugging. You will, too.</p><h3>The path that actually works: templates + guided iteration</h3><p>Premade templates are the future of vibe‑coding because they bake in architecture. Start from a strong, opinionated base; then iterate.<br>Example: use Nanobanana.biz to spin up a Nano Banana‑style AI image generator. You get:</p><p>A known‑good data model and file layout.<br>Core flows (upload, inference request, job queue, results gallery).<br>Guardrails around auth, billing, and rate limits.</p><p>From there, add features one at a time:</p><p>New model presets<br>Credits/limits<br>Webhooks for job completion<br>Moderation<br>Analytics</p><h3>A practical workflow I recommend</h3><p>Step 0: Pick a template that matches your domain (marketplace, SaaS, media, generator).<br>Step 1: Define your data model in plain language; have EasySite generate schemas and migrations; review them.<br>Step 2: Implement auth as its own task; test signup/login/verify/reset.<br>Step 3: Implement one core feature end‑to‑end (UI, API, DB, tests).<br>Step 4: Add observability (logs, error boundaries, basic metrics).<br>Step 5: Harden edges (pagination, empty states, retries, rate limits).<br>Step 6: Add one integration at a time (payments, email, storage). Test each in isolation.<br>Step 7: Refactor regularly; pay small debt early.</p><p>Red flags you’re drifting into Shit Code Mountain</p><p>The model keeps rewriting the same files in conflicting ways.<br>You can’t describe your entities and their relationships in two minutes.<br>Fixing a bug introduces two more in distant parts of the app.<br>You’re afraid to delete code because you don’t know what depends on it.</p><h3>Bottom line</h3><p>“One mega prompt to ship a complex product” is a fantasy. It sells because it’s easy to believe, not because it works.<br>AI is a fast coder but a sloppy architect unless you provide structure.<br>You win by planning data, slicing features, reviewing diffs, testing early, rolling back often.<br>Templates like Nanobanana.biz give you the architecture head start you need; then you iterate patiently.</p><p>If you want, tell me the app you’re building, and I’ll help you:</p><p>Pick a suitable template<br>Define the initial data model<br>Write the first 3–5 precise EasySite tasks to get to a working MVP without the mountain of junk code.</p><p>Here is a Pre-made Production Ready Template for Cloing Nano Banana:</p><p><a href="https://easysite.ai/nanobanana/landing-page">EasySite AI: Clone Any Website Instantly with AI - Enter URL &amp; Go</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ddf685cbad7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI Strikes Back: GPT-5-Codex and the Shift in the Coding Agent Wars]]></title>
            <link>https://medium.com/@luanxr2000/openai-strikes-back-gpt-5-codex-and-the-shift-in-the-coding-agent-wars-f3fd776bbc59?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3fd776bbc59</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Wed, 17 Sep 2025 10:25:57 GMT</pubDate>
            <atom:updated>2025-09-17T10:25:57.306Z</atom:updated>
            <content:encoded><![CDATA[<p>On September 16, OpenAI officially launched a new model, GPT-5-Codex — a fine‑tuned variant of GPT‑5 specifically designed for its various AI‑assisted programming tools. The company says the new model’s “thinking” time is more dynamic than earlier versions: the time it takes to complete a coding task can range from a few seconds to seven hours. As a result, it performs better on agent coding benchmark tests.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/603/1*oBU4yPeEj_g7dchHboT23w.jpeg" /></figure><p>The release of GPT‑5-Codex effectively brings to a close what may have been the most dramatic sentiment swing lately in the “coding agents” space.</p><p>For more than a year — starting from Claude 3.5 Sonnet last June, then 3.7 Sonnet and Claude Code in February, and Claude 4 in May — Anthropic had been far ahead in coding scenarios, holding a dominant position. During this period the company’s revenue skyrocketed to 5 billion USD (10% from Claude Code), and its market cap surged to 183 billion USD — an increase of 122 billion. All this clearly reignited OpenAI’s fighting spirit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*IjsWI6AT67q2uoGmO6apGQ.png" /></figure><p>Keep in mind, as early as 2021 OpenAI released the original Codex, which gave rise to GitHub Copilot — the world’s first AI programming tool (to which 182 developers are still actively contributing today). GPT‑3 also inspired Debuild, foreshadowing the later “vibe coding” startup wave.</p><p>Since then OpenAI has re‑prioritized coding capabilities in o1 and GPT‑4.1. GPT‑5-Codex scores 74.5% on SWE-bench, almost matching GPT‑5 (thinking) at 74.9% on the 477 subset.</p><h4>So what has driven the sharp turnaround in overall perception of GPT‑5?</h4><h4>One reason: the Codex team is “working extremely hard.”</h4><p><strong>First: a “multi-faceted unified” agent.</strong></p><p>Greg mentioned in a podcast today: “At the start of the year we set a company goal: build an agent-like software engineer by year’s end. Figuring out what that really means, how to implement it, and how to integrate all the opportunities and compute — this has been a huge task shared by many people at OpenAI.” The initial agent-style SWE shell was called 10X and ran in the terminal.</p><p>Now, with the new Codex CLI, “ChatGPT Codex” (now renamed Codex Cloud), IDE extensions (surpassing 800K installs in 2.5 weeks), and a GitHub code review bot, OpenAI has formed a complete set of interaction surfaces covering diverse needs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/796/1*cm23MWqXAFYxWTo62xhO3g.jpeg" /></figure><p><strong>Second: better post-training characteristics.</strong></p><p>OpenAI has always emphasized tight integration of research and product. The podcast also mentioned several important traits, the most significant being major improvements in “long-running agent tasks.”</p><p>Thibault Sottiaux said:</p><p>“This model shows a capability: it can persist longer, with the ‘grit’ needed for complex refactoring tasks. But at the same time, for simple tasks it responds very fast — it doesn’t overthink before giving an answer. That makes it a great collaborator — you can ask questions, locate code, plan a solution; and once you let it go, it can work continuously for a long stretch. We’ve seen it internally work seven hours straight to finish a complex refactor — something we’d never seen before. We’ve also invested heavily in code quality; it’s optimized for the real needs of Codex users.”</p><p>This cleverly leveraged “grit” is the key that makes GPT‑5-Codex a more comprehensive, practical agent-style programming model. It’s not just optimized for the hardest problems while forcing users to switch to a “dumber” model for easier tasks.</p><p>We have translated the full podcast interview so you can dive into how the OpenAI team built GPT‑5-Codex and the technology and stories behind it.</p><ol><li><strong>Why programming is a special exception in AGI research</strong></li></ol><p>Andrew Mayne:</p><p>Today we’re talking about Codex. I actually used it from its earliest versions back when I worked here. Now you have a new version — I spent the whole weekend with it. I’m blown away; I really didn’t expect the technology to advance this much in just a few years. I’d love to hear the origin story: how did the idea of using language models to write code first come about?</p><p>Greg Brockman:</p><p>I remember back in the GPT‑3 era, seeing the first signs: give it a docstring or a Python function definition and the model could complete the code. The first time you see that, you know it’s going to work — and it’s going to be big. We also talked about some idealistic goals at the time, like: imagine if a language model could write a thousand lines of coherent code — that would be amazing. That was our big target. Now that goal has long since been achieved and surpassed. These days we’ve grown used to it; it no longer feels shocking. But during development you mostly see the model’s gaps and flaws. Occasionally you step back and realize the technology has come incredibly far.</p><p>Thibault Sottiaux:</p><p>Yeah, what’s unbelievable is how quickly humans adapt to continuous improvement — it becomes an everyday tool almost instantly. You use it daily, then think back: a month ago this was impossible. It’s fascinating how fast humans adapt to novelty.</p><p>Greg Brockman:</p><p>We always had a dilemma — whether to go deep in a specific domain. Our mission is AGI, general intelligence. Intuitively, we want to raise all capabilities roughly equally. But programming has always been an exception. We had a completely different research program for it, focusing on programming data, code metrics, and model performance on coding tasks. Later we explored applying that approach to other areas too. But for programming, we consistently gave it special attention.</p><p>For GPT‑4, we ultimately released a general-purpose large model, but we also trained Codex models and Python‑leaning models. Around 2021 we pushed code capability to the max. I think the Codex demos we did then were maybe the earliest prototypes of what people now call vibe coding. I remember the interactive interface we built — it made me realize: ordinary language model interaction is simple; you complete a sentence or a chat. Code is different: it has to “live,” execute, connect to tools.</p><p>At that point you realize the “interaction harness” itself is as important as the intelligence model; it determines whether the model is truly usable. We understood that from that moment. This year, with stronger models, it’s not just about entering programming competitions or chasing raw capability — we want it truly usable. So we introduced diverse environments during training, linking the model to real development contexts plus suitable harnesses. That’s exactly what Thibault and his team have been pushing hard on.</p><p>Andrew Mayne: Can you explain “harness” in simpler terms?</p><p>Thibault Sottiaux:</p><p>It’s actually simple. The model itself is just an input‑output system. A harness is what integrates it with other infrastructure so it can actually operate in an environment. That includes tools and looping patterns — what we call the “agent loop.” Individually these pieces seem simple, but when you combine them end‑to‑end during training, you see magical behavior — the model can act on your behalf, create, and become a true collaborator. You can think of it like: the brain is the model; the harness is the body.</p><p>Andrew Mayne:</p><p>Exactly — very interesting. Think about the GPT‑3 days: we had to write commented code, like adding a # comment before a Python function to tell the model what it did. Now the model writes code naturally and intuitively. You mentioned general models vs. programming‑specialized models — is that driven by user demand or by your own desire to use it that way?</p><p>Greg Brockman:</p><p>Both. In 2022, we launched Copilot with GitHub. That’s when we first truly felt AI integrated into the programming workflow: it could speed you up. But there were lots of questions — like UI design: should it be “ghost text” autocompletion, or a dropdown of choices? One thing was clear: latency itself is a product feature. The threshold for autocompletion latency is 1500 ms — beyond that people won’t wait, no matter how smart it is. So the consensus at the time was: within latency constraints, use the smartest model you can. Then GPT‑4 arrived — it was smarter but didn’t meet the latency requirement. What do you do? We found the answer is: change the harness, change the interface. Interaction mode and model capability have to coevolve. A model that’s both fast and smart is great — but even if it’s slower, higher intelligence is worth it because intelligence’s payoff compounds long-term.</p><p>Andrew Mayne:</p><p>Back when we built GitHub Copilot, I didn’t grasp that point. I thought: as long as the model can complete code, it’s fine — didn’t realize how much the harness and tools matter. Now with the CLI, like Codex CLI, I can use it in the terminal; there’s a VS Code plugin; I can even deploy to the web. Back then I didn’t fully see the value. How do you personally use these things? Where do they feel most useful?</p><p>Thibault Sottiaux:</p><p>Returning to an initial observation: many developers were using ChatGPT to debug very complex issues. They’d keep stuffing more context in — code snippets, stack traces — then paste it to the model for help. As the interactions grew more complex, we suddenly realized: instead of making the user orchestrate the interaction, let the model fetch context, reason, and debug itself — then the user can just watch it work. That mental shift made us think more seriously about the harness and giving the model autonomy.</p><p>2. <strong>CLI, IDE, and terminal: different tools for different scenarios</strong></p><p>Greg Brockman:</p><p>We experimented with different forms. Early this year we tried multiple implementations: an async agent harness, local experiences, etc.</p><p>Thibault Sottiaux:</p><p>We also prototyped in the terminal. Later we felt that wasn’t “AGI‑enough” — we wanted it to scale, run remotely. You could close your laptop and it keeps working; you can check in from your phone. That felt so cool we pushed in that direction. But the terminal version was fully usable; many people at OpenAI were using it effectively. That tool was called “10X” at the time, because it really gave a 10x productivity boost. We didn’t productize it publicly because it didn’t feel polished enough. So we explored more shapes, first focusing on async. Now we’re circling back, bringing the agent into the terminal and IDE. The end goal is to make it a collaborator that lives inside the tools you already use.</p><p>Greg Brockman:</p><p>We tried other approaches too, like a remote daemon connected to a local agent — having the best of both worlds. We realized there’s almost a “form factor matrix”: async, local sync, hybrid. The question is where to focus: build a generalized external tool adaptable to any environment, or first nail the internal environment so our engineers have a stellar experience? We want both. But if we can’t use it well ourselves, how could the world? That’s the challenge — finding the right focus so engineering effort yields maximal value. This year we set a company goal: build an “agent-style software engineer” by year’s end. What does that mean? How do we implement it? How do we allocate resources and compute? Huge coordinated effort.</p><p>Andrew Mayne:</p><p>You mentioned “10X” — it looked extremely useful, yet you didn’t release it. Those decisions must be hard: when to go public and when not to. Cloud code is now powerful; I guess it’s a similar story — internal first, then external. How do you decide what to prioritize next? For me, the new form factor is so different it’s hard to fully grasp. Sometimes these things need time before people realize the value. Did your team have “aha moments”?</p><p>Greg Brockman:</p><p>Definitely. I think we can broadly sketch the future: you’ll want an AI with its own computer that can call a set of agents to parallelize tasks. In the morning, drinking coffee, you give it quick feedback like “This result is wrong” or “Adjust the process.” But models aren’t yet smart enough for that to be daily reality. So today, the model still needs to be in your terminal and editor — close to how you already work. That’s the present. For us, present and future are intertwined:</p><p>On one hand, bring AI into code review, proactively generating helpful content.<br> On the other, more PRs create new challenges — like how to filter what’s worth merging.<br> These opportunities are right in front of us. We’re seeing OpenAI’s internal development patterns and codebase structure begin to shift.</p><p>Thibault Sottiaux:</p><p>Yes — two key points behind the scenes. First: infrastructure is genuinely hard. We want code, tasks, packages perfectly containerized so we can run them at scale. Reality: many people have complex setups that only run on their own laptop. We want to meet users where they are so they don’t need custom setup just for Codex.</p><p>That lowers friction so people can immediately feel what a powerful coding agent can do. It also lets us experiment with what the “right interaction UI” is. Six months ago we weren’t even playing with these types of tools. Everything is new and evolving fast. We have to iterate relentlessly — what’s the right interface? The right collaboration pattern? Honestly, we don’t think it’s “done” — it will keep evolving.</p><p>But making it zero‑config, out‑of‑the‑box, extremely easy unlocks adoption and feedback — crucial for continued innovation.</p><p>Greg Brockman:</p><p>I remember early this year talking with one of our strongest engineers. He said ChatGPT had an integration that automatically saw terminal context. He said it was game-changing because he no longer had to copy‑paste errors. He could just ask “Hey, what’s wrong?” and it would tell him — worked amazingly. You realize this huge improvement came from integration itself, not solely from a smarter model.</p><p>It’s easy to latch onto one dimension and ask “Which matters more?” The answer: both. Back in 2020 when we first released the API, I thought of an AI’s “desirability” along two axes. One: intelligence (call it the vertical axis). Two: convenience (latency, cost, integration). There’s an “acceptable region” between them. A very, very smart model that takes a month to run — you might still use it if it produces extremely valuable code or, say, a cure. Conversely, if a model is less smart with limited capability, you only want it for autocomplete — then it must be ultra-convenient — minimal cognitive overhead.</p><p>Right now we sit at some point on that spectrum: smarter models that aren’t as convenient as instant autocomplete, but vastly more convenient than “wait a month.” One challenge: deciding when to invest in pulling convenience left (lower latency, lower cost, better integration) versus pushing intelligence up. Huge design space — that’s what makes it interesting.</p><p>Andrew Mayne:</p><p>Right. I made an app featured at the GPT‑3 launch called AI Channels. The challenge: GPT‑3 was powerful, but I had to write a 600‑word prompt to get it to work. At $0.06 per 1K tokens plus latency, I felt: this world isn’t ready yet. Later GPT‑3.5 and GPT‑4 arrived — you suddenly saw capabilities click. Hard to articulate why, but everything aligned.</p><p>You also mentioned letting the model see context where you work. I used to copy from ChatGPT into my workspace — like entering a supermarket refusing a cart, carrying everything by hand — deeply inefficient. Once you put stuff in a wheeled cart, it all flows. We’re seeing those “unlock” moments now. My current dilemma: when I sit down to work, do I go into the CLI? Use the VS Code plugin? Cursor? Something else? How do you choose?</p><p>Thibault Sottiaux:</p><p>Right now we’re still experimenting — trying different ways to let you interact with the agent inside the places you’re already productive. For example, Codex is now in GitHub. You can @Codex and it will get to work. If you write “@Codex fix this bug” or “move the tests here,” it spins up and executes the tasks on its own “little notebook” (dedicated compute) in our data center — you don’t worry about it. But if you’re manipulating a folder and files, you face a choice: in the IDE or in the terminal?</p><p>We’ve observed that advanced users develop complex workflows in the terminal; but when you’re focused on a specific file or project, you prefer the IDE because the interface is richer: undo, visual edits — not just scrolling output.</p><p>Meanwhile, the terminal is a great “vibe coding” tool: if you don’t care as much about the precise code text, you can conjure a small app in place. That’s more about interaction and results rather than the textual artifact. So it depends on what you’re doing. We’re still in the experimental phase, and I think it will continue for a while.</p><p>Greg Brockman:</p><p>I fully agree. I also think our future direction is stronger unification. Developers already use multiple tools: terminal, browser, GitHub web UI, local repos. Each has learned rules for “when to use what.” Because we’re still experimenting, these things look fragmented — you feel you have to learn a new set of affordances for each.</p><p>As we iterate, we must think carefully about how they fit together. You can already see trends: the Codex IDE extension can run remote Codex tasks. Ultimately we envision an AI that can both use its own computer/cluster and “look over your shoulder” locally. Those shouldn’t be separate worlds.</p><p>Thibault Sottiaux:</p><p>Yes — we think of it as a “single coding entity” that’s always there helping and collaborating. Like me working with Greg — sometimes Slack, sometimes in person, sometimes via GitHub reviews — it’s natural. Our vision for Codex is the same: an agentic entity that truly “accelerates” you in achieving goals.</p><p><strong>3. The limits of agent memory</strong></p><p>Andrew Mayne: Let’s talk about usage patterns — like agents.md. Can you explain it?</p><p>Thibault Sottiaux:</p><p>Sure. agents.md is an instruction file you can provide to Codex stored alongside your code. That gives Codex additional context so it can navigate the codebase and accomplish tasks better.</p><p>We found two categories of content are especially useful there:</p><p>(1) “compressed” knowledge — so the agent can read codex.md (original wording) efficiently instead of exploring the entire codebase exhaustively; (2) “preferences” not explicitly encoded in the code itself — for example: where tests should really go, or “I prefer doing X in this particular style.” These two — preferences and guidance for efficient navigation — are ideal for agents.md.</p><p>Greg Brockman:</p><p>There’s a fundamental issue: how do you communicate “what you want and your preferences” to a “stateless” agent and also save it the human warm‑up phase? For humans we write readme.md. agents.md is just a conventional filename telling the agent “go look there.”</p><p>But it also reflects a “moment-in-time reality”: current agent memory isn’t great. For example, on your tenth run, is it really benefiting from the prior nine times it helped you with hard problems? So we do have research ahead: how to give it memory? How to let an agent really explore your codebase, deeply internalize it, then leverage that knowledge? This is just one among many examples, but we already see lots of low‑hanging research fruit ahead</p><p><strong>4. Migration and review</strong></p><p>Andrew Mayne: Competition is intense now. There was a time when OpenAI felt like it appeared out of nowhere — suddenly GPT‑3, then GPT‑4. Now Anthropic has great models, Google’s Gemini is strong. How do you view the landscape? Where do you see yourselves?</p><p>Greg Brockman:</p><p>There’s still huge room to improve. I focus less on competition and more on potential. When we founded OpenAI in 2015, we believed AGI would arrive sooner than most expected. We just wanted to play a positive role. Translating that into execution has always been core. When we explored how to build truly useful models, getting them into people’s hands became critical. Look at choices we made — launching ChatGPT, broadly free access. Mission-driven: we want AI broadly accessible and beneficial. To me, the key is sustaining exponential progress while always asking how to deliver it positively and usefully.</p><p>For example, we have pre‑trained GPT‑4 class models, then layer reinforcement learning to make them more reliable and intelligent. Imagine: if you only “read the Internet,” passively observing fragments of human thought, then wrote code for the first time — you’d be awful.</p><p>But if you could actively solve tough programming problems with a Python interpreter and human-like tools, you’d become more robust and refined. Now we combine those elements — but must push to the next tier.</p><p>Large‑scale codebase refactors — no one has truly cracked that yet. There’s no fundamental barrier. Once solved, I think it becomes a “killer app” for enterprises. If refactor/migration cost halves, migration volume may 10x. Consider COBOL — huge systems trapped there, but no new COBOL programmers. That’s a global risk/liability. The only way out is systems that can truly tackle it. Massive open space. Exponential curves will continue, and we have to keep pace.</p><p>Andrew Mayne:</p><p>My favorite little tidbit today was OpenAI posting a tweet teaching people how to switch the completions API to the responses API in the CLI.</p><p>Thibault Sottiaux:</p><p>That’s a great use case. I expect more like it — giving Codex special instructions so it reliably performs refactors and just runs them. It’s wonderful. Migration is one of the most painful chores. Nobody wants to migrate — from one library to another while ensuring everything still works. If we automate most of that, it’s a beautiful contribution.</p><p>Greg Brockman:</p><p>Tons of other directions too. Security patching will become vital — we’re investing there. Also, AI can generate new tools. Think of the importance of the Unix toolset — if AI can build tools for you, even for itself, it constructs an ever-growing ladder of complexity and utility, spinning an efficiency flywheel. Future AI won’t just write code; it will execute its code, manage services, even do SRE‑type work. Those are on the horizon — already beginning — but not yet at our ideal.</p><p>Thibault Sottiaux:</p><p>One big internal problem we solved at OpenAI is code review. As code volume grows, the bottleneck is the sheer review load. So we built a high‑signal Codex mode that can deeply review PRs, understand developer intent, then verify if code implements that intent. It can drill down through dependencies, think about contract logic, and surface issues that even top reviewers would take hours to find. We launched it internally first — super successful.</p><p>When it goes down, people get upset — they feel they lost a safety net. It massively accelerated teams, including Codex itself. The night before we shipped the IDE extension, one of my top engineers submitted 25 PRs. We automatically found quite a few bugs — Codex surfaced many issues. Next day we shipped with practically no bugs — astonishing speed.</p><p>Greg Brockman:</p><p>What’s interesting: people were initially nervous about a code review tool. Earlier experiments with automated review were “noise” — emails from a bot you just ignore. Now it’s the opposite. Below a certain capability threshold, it’s net negative; above it, people actively need it and complain when it’s gone. Our observation: if an AI feature is only “somewhat useful” today, a year later it will be extremely reliable and mission-critical. That’s the path code review is on.</p><p>Thibault Sottiaux:</p><p>Code review also has an interesting angle: bringing humans along — making it a collaborator. We spent a lot of thought on making the review something you want to read — even if it’s occasionally wrong. You can understand its reasoning. In over 90% of cases it’s correct. Even as author or reviewer, you often learn from it.</p><p>Greg Brockman: Returning to “pace of progress”: think of GPT‑3 vs. GPT‑4 — we were worried about “double down” behavior: AI says something wrong, you point it out, it argues.</p><p>Andrew Mayne: Right — it insists it’s correct.</p><p>Greg Brockman:</p><p>Exactly. We’re far past that stage. Sure, rare edge cases still happen — humans do too. But now we’re at a level where even if it doesn’t fully nail the point, it still highlights something valuable — a legitimate line of thought. Every time I read these reviews I think: “Yeah, that’s right — I should consider that.”</p><p><strong>5. GPT‑5 Codex’s “seven-hour miracle”</strong></p><p>Andrew Mayne: Tell us why we should be excited about GPT‑5 Codex.</p><p>Thibault Sottiaux:</p><p>GPT‑5 Codex is a version of GPT‑5 we optimized for Codex. We mentioned the harness — it’s optimized for that. We treat it as a single agent tightly integrating model and tools — thus more reliable. A standout trait: it can persist longer, with the “grit” for complex refactors. Simultaneously, for simple tasks it’s fast — no overthinking before replying.</p><p>It’s a great collaborator: you can ask code questions, have it locate changes, help you understand and plan. Once you let it run a task, it can continue for a long stretch. Internally we’ve seen it work seven hours continuously on a complex refactor — unprecedented for other models. We’ve also invested heavily in code quality. It’s optimally tuned for Codex usage scenarios.</p><p>Andrew Mayne: When you say it can persist seven hours, it’s not just stuffing context back in — it’s actually making decisions, judging importance, moving forward?</p><p>Thibault Sottiaux: Yes. Imagine a particularly thorny refactor. We’ve all been there: codebase is no longer maintainable; you must restructure to keep going. You draft a plan, hand it to GPT‑5 Codex; it resolves issues step by step, gets tests passing, and completes the refactor. That’s what we saw in that seven-hour continuous case.</p><p>Greg Brockman: What amazes me is how strong the core intelligence already is. Three to six months ago these models were already better than me at finding specific functionality inside internal codebases. That requires fairly complex ability.</p><p>Andrew Mayne: So are you “out of a job,” like “Greg, we’re sorry…”</p><p>Greg Brockman:</p><p>(laughs) The truth is I can do more other things. I don’t want to be known for “searching code.” That’s not how I define my engineering value. I want to think about architecture, decide how many agents to run, what tasks to allocate. That’s the core value: AI handles the dull mechanics and also partners on the interesting parts. I can choose how to allocate my attention.</p><p>I’m a die-hard Emacs user, but I also use VS Code, Cursor, Windsurf — partly to try them, partly because I enjoy tool diversity. But it’s hard to drag me out of a terminal. Now we’re at the threshold where during a refactor I think: “Why am I typing this manually? Why recall some syntax?” It’s mechanical. I’d like an intern to do it — and now I have that intern in my terminal. It’s amazing.</p><p>Andrew Mayne:</p><p>Let’s talk about the future. For agentic futures, what’s most exciting? How will we use these systems?</p><p>Thibault Sottiaux:</p><p>We’re very certain the future form is: large populations of cloud agents, supervised and guided by humans, teams, organizations — creating huge economic value. In the coming years that’s the picture: millions of agents doing useful work in organizational data centers.</p><p>The question: how do we get there step by step? What’s the right shape, interaction mode? One major problem to solve is safety, safeguards, alignment. Agents must do useful work — but safely. You must stay in control as operator. That’s why in Codex CLI, by default, the agent runs in a sandbox; it can’t arbitrarily modify files on your computer. We’ll keep investing in safer environments, understanding when human approval is needed, gradually granting more permissions. Your agent will have a set of permissions you allow; it can “elevate” with your consent for higher-risk actions.</p><p>Thibault Sottiaux:</p><p>So you understand the whole system, then make it multi-agent so individuals, teams, organizations can guide it while remaining aligned with organizational goals — that’s the direction. Abstract, but very exciting.</p><p>Greg Brockman:</p><p>I fully agree. On a more granular level there are many technical problems: “scalable oversight” that Thibault mentioned — how do humans manage agents that keep producing lots of code? You probably won’t read every line. Most people don’t now. How do you preserve trust? Ensure correctness? There are technical avenues. Since 2017 we’ve explored strategies for weaker AIs or humans supervising stronger AIs, bootstrapping trust even for high-difficulty tasks. Vital question — very concrete for advanced coding agents.</p><p>Another often overlooked dimension: at each capability tier people overfit, thinking “Ah, this is AI; this is how it’ll always be.” We haven’t truly seen AI solving genuinely new, hard problems yet. Today people think “I need a refactor” — you know roughly what that looks like; AI saves time. But what about problems unsolvable by other means? I’m thinking beyond programming — medicine (new drug discovery), materials science (novel properties).</p><p>Soon new capabilities will unlock those. A key milestone: AI outputs with intrinsic value — not because they’re cheap or “done by AI,” but because they’re breakthroughs — even if achieved through human collaboration.</p><p>We already see hints. In life sciences, researchers had o3 generate five experiment plans — four failed, one succeeded. Feedback: about a third/fourth-year PhD student level — astonishing, and that was o3. With GPT‑5 and GPT‑5 Pro we see a different landscape. Scientists say: “This is genuine novel research.” Sometimes it’s not fully cracking a major theory alone, but extending human reach. That’s what we must keep pushing — done right.</p><p><strong>6. The future of software engineering</strong></p><p>Andrew Mayne: A challenge when discussing the future: people imagine it as “the present with shiny coating” — robots everywhere, or robots write all code. But you said some things you like doing, some you don’t. What does 2030 look like? Five years ago it was GPT‑3 — so five years ahead?</p><p>Thibault Sottiaux: Six months ago we didn’t even have these tools, so picturing five years out is hard. One important point: the critical infrastructure code that underpins society — we must keep understanding it and have tools to do so. That’s why we value code review — code review should help you understand code, be a teammate, help you dive into others’ code — including AI’s code.</p><p>Greg Brockman: We already have a problem: lots of code “out there” isn’t safe. Happens frequently. I remember Heartbleed twelve years ago — a severe vulnerability in core widely used Internet software. And you realize it’s not an isolated case — there are many similar undiscovered flaws.</p><p>Andrew Mayne: Right, plus so many NPM packages — some abandoned — people could slip malicious code in.</p><p>Greg Brockman:</p><p>Exactly. Historically it’s attacker‑defender arms race — each escalating. With AI you ask: which side gains more? Maybe it accelerates the cycle. But there’s hope AI unlocks entirely new defense capabilities. Formal verification could be a defense “endgame.” What excites me is not just continuing a grindy arms race, but producing higher stability and comprehensibility. Other areas will mirror this — giving us novel ways to understand systems. Many traditional software systems are already pushing human comprehension limits.</p><p>Thibault Sottiaux:</p><p>One reason we built Codex is to improve infrastructure and existing code — not simply increase total lines of code. That matters. Codex can find bugs, refactor, discover more elegant, efficient implementations — achieving the same goals (or more general ones) without spawning 100 million unintelligible lines. I’m excited that Codex can help teams and individuals write better code, become better engineers, and ultimately produce leaner systems that do more.</p><p>Greg Brockman:</p><p>I think a key 2030 picture: a world of “material abundance.” AI will make creating what you want far easier — not just digitally, physically too. But at the same time: “extreme compute scarcity.” Internally we already feel it — projects jockey for compute; research success hinges on allocation — hard to overstate. I think the future’s similar: creativity constrained not just by imagination, but by compute. So we think a lot about increasing global compute supply. We must raise intelligence and its availability. That’s fundamentally a physical infrastructure issue, not just software.</p><p>Thibault Sottiaux:</p><p>What amazes me about GPT‑5 is we can offer it in Free, Plus, and Pro tiers. Subscribe to Plus and you get Codex and GPT‑5 — the same version others use — highly efficient at that access level.</p><p>Andrew Mayne: I’m struck that this model is more powerful yet priced the same or lower. That shocked me.</p><p>Greg Brockman:</p><p>Yes. We’re increasing intelligence while dramatically lowering price. Easy to overlook — pretty wild. For example, we cut o3 costs by 80%. Back in GPT‑3 days it was $0.06 per 1K tokens.</p><p>Andrew Mayne:</p><p>Yeah, there was a recent report complaining “reasoning models” are pricier — but they didn’t compare them to new reasoning models over the last 6–7 months or note efficiency gains.</p><p>Greg Brockman:</p><p>Exactly. And the trend continues. On compute scarcity: a clear sign — people discuss million‑GPU clusters now. Soon maybe everyone needs a personal GPU for their agent. That implies nearly 10 billion GPUs. We’re orders of magnitude away. Part of our job: expand supply while extracting maximal value from current limited compute — a dual efficiency and intelligence push. Achieving it requires huge building effort.</p><p>Thibault Sottiaux:</p><p>Another interesting piece: agent–GPU proximity. If the GPU is closer to the user there’s a huge benefit. An agent might call tools 200 times over minutes — each round trip needs GPU ↔ laptop interaction. If the GPU is closer, latency drops; interaction loop smooths dramatically.</p><p>Andrew Mayne:</p><p>People often ask about the future and the workforce: should they learn programming?</p><p>Thibault Sottiaux: I think now is a fantastic time to learn programming.</p><p>Greg Brockman: I agree. Definitely learn to program — but more importantly, learn to use AI. That’s most crucial.</p><p>Thibault Sottiaux: Using Codex to learn a new language is fun. Several on our team didn’t know Rust, yet we decided to write the core harness in Rust. They used Codex to ask questions, explore unfamiliar codebases, ramped up quickly, and delivered great results. Of course we also had experienced Rust engineers to ensure quality. Overall, it’s the best time to learn.</p><p>Greg Brockman:</p><p>When I first learned programming I used W3Schools tutorials — PHP, JavaScript, HTML, CSS. Building my first app I wanted to serialize data, but didn’t know the word “serialize.” I invented a custom delimiter sequence. But what if the data itself contained that sequence? Disaster (laughs). Tutorials didn’t warn me. Codex in code review would just say: “Hey, there’s a JSON serialization library — use that.” So barriers drop — you don’t reinvent wheels; it surfaces questions you never thought to ask and answers them. That’s why I believe now is the best era to build things.</p><p>Andrew Mayne:</p><p>I also learn by watching it solve problems — discover new libraries, new approaches. Sometimes I give it wild tasks like: build a language model in 1,000 lines of code — what would it do? Sometimes it fails, but you see its attempts and think “Ah, that’s another angle.”</p><p>Thibault Sottiaux:</p><p>I’ve noticed the people who leverage AI best typically have solid software engineering fundamentals. They design reasonable architecture, scaffold code, then let AI help — thus they go further. If you know nothing about code you can’t reach that stage.</p><p>Andrew Mayne: Since you launched GPT‑5 and Codex deployability, how’s usage?</p><p>Thibault Sottiaux: Explosive growth. Overall usage up over 10x. Existing users use it more often, in more complex scenarios, for longer durations. We’ve integrated it into Plus and Pro with generous quotas — also a driver.<br> Greg Brockman: The “vibe” is shifting. People are realizing the optimal way to use GPT‑5. We have our own interaction harness and tool combination. Once users internalize it their velocity spikes.</p><p>Andrew Mayne: Thanks so much to both of you for the conversation. Any final thoughts?</p><p>Greg Brockman: Thanks for having us! We’re really excited about what’s ahead. Lots more to build. Exponential progress continues — and making these tools broadly accessible and truly useful is our mission.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3fd776bbc59" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Integrating OpenAI’s GPT Realtime Model with Twilio]]></title>
            <link>https://medium.com/@luanxr2000/integrating-openais-gpt-realtime-model-with-twilio-f4039d85c3bd?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4039d85c3bd</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Fri, 12 Sep 2025 04:40:07 GMT</pubDate>
            <atom:updated>2025-09-12T04:40:07.277Z</atom:updated>
            <content:encoded><![CDATA[<h3>1.0 Executive Summary: The Conversational AI Phone Agent Blueprint</h3><p>This report provides a comprehensive architectural blueprint and implementation guide for building a real-time AI phone agent by integrating OpenAI’s GPT Realtime API with the Twilio Voice platform. This solution represents a significant advancement over traditional Interactive Voice Response (IVR) systems, offering a low-latency, “speech in, speech out” conversational experience that feels remarkably natural and human-like.1 The strategic value of this technology lies in its ability to automate a wide range of customer interactions at scale, from fielding general inquiries to complex tasks like appointment scheduling and lead qualification.1 By leveraging the global telecommunications infrastructure provided by Twilio and the advanced conversational intelligence of OpenAI’s models, businesses can deliver personalized customer experiences with 24/7 availability and without the fatigue associated with human agents.1</p><p>The core architectural recommendation presented herein is centered on a custom application server that functions as a high-performance, asynchronous WebSocket bridge. This server intelligently orchestrates the real-time audio and event streams between Twilio and OpenAI, ensuring a seamless, two-way conversation. The report details the critical role of Twilio’s bidirectional Media Streams (&lt;Connect&gt;&lt;Stream&gt;) and the WebSocket protocol, which together form the foundational communication layer.4 The central innovation, however, is not the API integration itself but the sophisticated server-side logic required to manage conversational state, minimize perceived latency, and, most critically, implement a robust system for handling interruptions and preemption. This is the key technical feature that transforms a clunky, turn-based bot into a genuinely responsive conversational partner.5</p><p>The commercial success of a platform like NewOaks AI serves as a powerful case study for this blueprint. The product’s advertised capabilities — such as a 24/7 AI receptionist, automated appointment booking, and outbound lead qualification — are direct results of a mature implementation of this architectural model.1 NewOaks AI also exemplifies a broader business trend: the monetization of a modular AI backend that can be generalized across multiple communication channels, including web chat and SMS, in addition to voice.6 This approach allows the technical core to serve as the foundation for a suite of “AI-as-a-Service” products, moving beyond a single-function solution. Ultimately, a successful and compliant deployment requires not only technical proficiency but also a deep understanding of security best practices, scalability considerations, and the regulatory landscape for automated calling, such as the Telephone Consumer Protection Act (TCPA).8</p><h3>2.0 Introduction: From Voice to Value, a New Era of Customer Interaction</h3><h3>2.1 The Problem: The Voice Interaction Gap</h3><p>Modern telephony has struggled to keep pace with the advancements in digital communication. For many businesses, the customer service experience remains a point of friction, defined by frustrating, menu-driven Interactive Voice Response (IVR) systems that fail to resolve complex queries and often force callers into long, unproductive queues.3 For businesses, the alternative — maintaining a large, human-led contact center — is a significant operational cost that often prevents scaling customer support to a global, 24/7 presence.1 The challenge is to bridge this gap, offering the immediacy and ubiquity of a phone call with the efficiency and intelligence of modern AI.</p><h3>2.2 The Solution: The Twilio and OpenAI Synergy</h3><p>A new paradigm has emerged, enabled by the convergence of two key technologies: Twilio’s global voice network and OpenAI’s GPT Realtime API. Twilio serves as the foundational telecommunications layer, providing a programmable platform to make, receive, and manage phone calls at a global scale.10 Its APIs give developers precise control over call flow using a simple, declarative language called TwiML (Twilio Markup Language).4</p><p>OpenAI, in turn, provides the conversational intelligence engine. The GPT Realtime API is a specialized interface for the gpt-4o-realtime-preview and gpt-realtime model families, designed explicitly for low-latency, “speech-in, speech-out” interactions.2 Unlike standard API calls that require the entire audio file to be uploaded before a response can be generated, the Realtime API operates over a WebSocket connection, allowing for a continuous, bidirectional stream of audio data.2 This architecture is essential for achieving a truly natural, conversational cadence. The synergy between these two platforms is made possible by Twilio’s Media Streams, which allows the raw audio from a live phone call to be forked and sent to a WebSocket server in near real-time.4</p><h3>2.3 The Case Study: NewOaks AI as a Reference Implementation</h3><p>The NewOaks AI platform provides a powerful, real-world example of this technology stack in a commercial setting.1 Its core features — including a 24/7 AI receptionist, automated appointment booking, and outbound lead qualification — are not just hypothetical capabilities but are demonstrated applications of the Twilio and OpenAI integration.1 The platform’s ability to provide human-like voice agents that are always available and can perform complex tasks like scheduling appointments with a live calendar underscores the potential of this architecture.1</p><p>An analysis of NewOaks AI’s product suite reveals a broader strategic pattern beyond just phone agents. The company also offers AI-powered website builders, chatbots, and SMS services, and its SEO strategy reflects competition with broader AI platforms like Durable.co, Wix, and WordPress.12 This suggests that the technical blueprint for the real-time phone agent is a foundational component of a general-purpose, “AI-as-a-Service” business model. The core technology — integrating an LLM with a communication channel — is modular and can be extended to various platforms, from voice calls to web chat and text messaging.6 The ability to create a consistent, AI-driven experience across multiple touchpoints demonstrates a sophisticated understanding of the market, where the underlying AI engine is the primary value proposition, and communication channels are simply different delivery mechanisms.6</p><h3>3.0 Core Technology Deep Dive: The Real-Time Voice Pipeline</h3><h3>3.1 Twilio’s Role: The Telecommunications Backbone</h3><p>Twilio provides the robust and scalable infrastructure necessary to connect a conversational AI to the public switched telephone network (PSTN). Its Programmable Voice API is controlled through TwiML, an XML dialect used to provide instructions for a call. The key TwiML verb for this architecture is &lt;Connect&gt;&lt;Stream&gt;, which initiates a bidirectional audio stream between the live call and a specified WebSocket server.4 This stream is the essential conduit for two-way, real-time communication.</p><p>Twilio’s Media Streams functionality forks the raw audio from the call and sends it as a continuous stream of data to a WebSocket endpoint that the developer specifies.4 The stream is bidirectional, meaning the WebSocket server can both receive audio from the caller (</p><p>inbound_track) and send audio back to the caller (outbound_track). The audio is streamed in a compressed format that is lightweight enough for real-time transmission, eliminating the need for complex transcoding on the server side.8 This direct, low-overhead communication channel is what allows for the low-latency conversation required by AI agents. A critical security measure is the validation of the</p><p>X-Twilio-Signature header, which ensures that incoming streams are genuinely from a trusted Twilio source and have not been tampered with.4</p><h3>3.2 OpenAI’s Role: The Conversational Intelligence Engine</h3><p>The OpenAI GPT Realtime API is a specialized service designed to support low-latency, “speech-in, speech-out” interactions, a fundamental requirement for a fluid conversation.2 The API is an extension of the GPT-4o family of models and is optimized for conversational use cases. It supports both WebRTC and WebSocket protocols, making it compatible with a wide range of real-time applications.2 The API allows developers to provide the model with instructions and context, defining the AI assistant’s personality and specifying its behavior, such as what it should and should not answer.2 This is crucial for tailoring the agent to specific business use cases, such as acting as an AI receptionist or a lead qualifier.1 OpenAI’s Realtime API provides key events, such as</p><p>speech_started, which signal to the server when the user has begun speaking, a critical component for implementing a preemption or interrupt system.5</p><h3>3.3 The Architectural Nexus: A WebSocket Bridge</h3><p>A common misconception is that Twilio and OpenAI can connect directly. In reality, a custom application server is required to act as a bidirectional proxy. This server’s function is to receive the audio stream from Twilio’s WebSocket and, after processing, forward it to OpenAI’s WebSocket, and vice-versa. This middle layer is the central hub for all conversational logic, managing the data flow and orchestrating the user experience.</p><p>The core of this architecture is the handling of WebSocket events and the management of audio data. Twilio sends a stream of JSON-encoded messages that contain Base64-encoded audio payloads.13 The server must be programmed to parse these messages, decode the audio, and forward it in the appropriate format to OpenAI. Similarly, the server receives audio from OpenAI in real-time and must forward it to Twilio’s stream for playback to the caller.5</p><p>The success of a “real-time” agent is not merely about transmitting data quickly; it is about minimizing the perceived latency. This requires the server to be an active orchestrator, not a passive pipe. The server’s logic must include:</p><ul><li><strong>Chunk-based Processing:</strong> The server should not wait for the user to finish speaking before sending audio to the AI. It should process and stream audio in small chunks as they arrive to enable an almost immediate response.</li><li><strong>Interrupt Handling (Preemption):</strong> A truly natural conversation requires the ability to interrupt the other party. The server must monitor for Twilio’s real-time events that signal when the user has begun speaking and, upon detection, send a truncate message to OpenAI’s WebSocket to immediately stop the AI’s current response.5 This feature is what prevents the AI from talking over the user and is a primary differentiator between a natural conversation and a rigid, turn-based dialogue.9</li></ul><p>The following table provides a detailed breakdown of the WebSocket event flow that is the lifeblood of this architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/858/1*AllXpMVmKprHV6irihS8ow.png" /></figure><h3>4.0 Implementation Blueprint: A Step-by-Step Guide for Inbound and Outbound Calls</h3><h3>4.1 Prerequisites and Project Setup</h3><p>Before implementation, a developer requires the following:</p><ul><li>A Twilio account with a voice-enabled phone number.5</li><li>An OpenAI account with a valid API key and access to the Realtime API.2</li><li>Python 3.9+ or Node.js LTS.2</li><li>A local development environment configured with a virtual environment.8</li><li>The following libraries installed via pip or npm: fastapi, uvicorn, twilio, websockets, python-dotenv for Python; or fastify, ws, twilio, dotenv for Node.js.8</li><li>A local tunneling solution like ngrok to expose the development server to the public internet, as Twilio requires a public URL for its webhooks.5</li></ul><p>A .env file should be created to securely store sensitive credentials, such as API keys and phone numbers, as environment variables.8</p><h3>4.2 Handling Inbound Calls</h3><p>An inbound call is initiated when a caller dials the Twilio phone number. Twilio is configured to send a webhook request to a specified URL when a call comes in, typically a /incoming-call endpoint on the custom server.5 The server’s logic for this endpoint must respond with TwiML instructions that define the call flow.</p><p>The TwiML response typically first uses the &lt;Say&gt; verb to play a short greeting to the caller.9 Immediately following this greeting, the TwiML must include the</p><p>&lt;Connect&gt;&lt;Stream&gt; noun, which instructs Twilio to fork the audio stream to the server’s WebSocket endpoint.11 The</p><p>url attribute of the &lt;Stream&gt; noun must point to the public, ngrok-generated WebSocket URL, for example: wss://[your-ngrok-domain].ngrok.app/media-stream.11</p><h3>4.3 Handling Outbound Calls</h3><p>Outbound calls are initiated programmatically from the custom server using the Twilio REST API.8 The</p><p>client.calls.create method is used to create a new call. The API call requires a to parameter (the destination phone number) and a from_ parameter (the Twilio phone number to call from).8 The</p><p>twiml parameter is the most critical element, as it tells Twilio how to handle the call once it is connected.8 Similar to inbound calls, this</p><p>twiml specifies that the call should be connected to the WebSocket stream endpoint for the AI agent.8</p><p>For outbound calling, it is essential to incorporate compliance and number validation. The server-side logic should include a function to check if the destination number is an owned Twilio number or a verified outgoing caller ID.8 This is a critical step for adhering to Twilio’s terms of service and legal regulations, particularly those concerning automated calls.8</p><h3>4.4 The WebSocket Server Logic</h3><p>The core of the application resides in the WebSocket server logic, which handles the bidirectional communication. When a Twilio call connects to the /media-stream endpoint, the server must first accept the connection and then immediately initiate a new WebSocket connection to the OpenAI Realtime API.5</p><p>The server then enters a continuous event loop, listening for messages from both WebSockets.13 When a Twilio</p><p>media event is received, the server’s logic should:</p><ol><li>Parse the JSON message.13</li><li>Extract the Base64-encoded audio payload.13</li><li>Decode the payload to its raw audio format.13</li><li>Forward this raw audio chunk to the OpenAI WebSocket.5</li></ol><p>The server also manages the conversational state. For instance, an initial greeting can be sent to OpenAI, which will then generate the corresponding audio and stream it back to the server. The server forwards this audio to Twilio, which plays it to the caller.8 The most sophisticated part of the implementation is the interrupt handling. By monitoring events from Twilio and OpenAI, the server can detect when the user starts speaking and immediately send a</p><p>truncate command to OpenAI, clearing the audio buffer and stopping the AI from talking over the user.5</p><p>A closer look at the pricing model of a company like NewOaks AI reveals a crucial strategic decision: the “Bring Your Own Key” option.16 This is more than just a pricing tactic; it addresses a fundamental enterprise concern regarding data security and cost management. By allowing clients to use their own OpenAI API key, the platform sidesteps the responsibility of handling sensitive client data, as the data is processed directly within the client’s own OpenAI account.16 Additionally, this shifts the variable cost of API usage directly to the customer, allowing NewOaks AI to focus on charging a fixed subscription for its platform, integrations, and value-added features like the user interface and analytics.7 This strategic move transforms the business from a per-use model into a scalable platform-as-a-service model.</p><h3>5.0 Case Study: Deconstructing the NewOaks AI Solution</h3><p>NewOaks AI’s feature set is a prime example of how the foundational Twilio-OpenAI blueprint can be extended to create a full-fledged, commercial product. The platform’s advertised capabilities are direct manifestations of an advanced, multi-component technical solution.1 The following table maps NewOaks AI’s key features to the specific technical components required for their implementation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*r4WE8zHDdybwx1JIgIjiWA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d_RrRiq_9lvc_5WQuazqqg.png" /></figure><p>A truly conversational agent that can perform complex tasks, such as scheduling an appointment, requires more than just speech-to-speech capabilities. The platform must be able to act on the information it receives. The core enabling technology for this is the concept of tool use within the OpenAI model.17 By defining specific functions (e.g.,</p><p>create_calendar_event or update_crm_lead), the AI can recognize a user’s intent to perform a specific action and then call the corresponding function on the server.17 This server-side function then executes the real-world task by interacting with a third-party API, such as a calendar system or a CRM like HubSpot or Salesforce.6</p><p>This capability is what elevates a conversational AI from a novelty into a genuine business solution. A bot that can only answer questions is a chatbot; a bot that can actually book a meeting, update a customer’s record, or process a payment is a workflow automation tool.6 NewOaks AI’s reliance on CRM, calendar, and custom API integrations is a testament to this understanding, recognizing that the real value lies in the AI’s ability to seamlessly transition from conversation to action, a strategic imperative for any enterprise-grade solution.6</p><h3>6.0 Security, Scalability, and Compliance Considerations</h3><h3>6.1 Security Best Practices</h3><p>Implementing a solution that handles real-time voice data and sensitive API keys requires a rigorous focus on security. API keys for both Twilio and OpenAI should never be hard-coded directly into the source code.8 Instead, they should be stored securely as environment variables in a</p><p>.env file and loaded at runtime using a library like python-dotenv.8 For production deployment, a secrets management system should be used.</p><p>Furthermore, a critical security measure is the validation of incoming webhooks. The Twilio platform signs its webhook requests with a unique signature, which is passed in the X-Twilio-Signature header.4 The custom server should be configured to validate this signature to ensure that the request is genuinely from Twilio and not a malicious, spoofed request.4</p><h3>6.2 Scalability and Reliability</h3><p>The WebSocket-based architecture is inherently scalable due to its asynchronous nature. Frameworks like FastAPI (Python) and Fastify (Node.js) are built on asynchronous principles, allowing them to handle a large number of concurrent connections efficiently without blocking.8 For a production-grade deployment, a local tunneling solution like</p><p>ngrok must be replaced with a robust cloud-based hosting provider.8 Deploying the application server to a location close to the intended user base can also help minimize network latency and improve the conversational experience.4</p><h3>6.3 Compliance and Ethical Considerations</h3><p>The deployment of automated voice agents requires careful consideration of legal and ethical guidelines, particularly for outbound calls.8 In the United States, the Telephone Consumer Protection Act (TCPA) regulates automated phone calls, and compliance is mandatory.8 A crucial best practice, and often a legal requirement, is to disclose to the caller that they are speaking with an AI.8</p><p>A fundamental tension exists between the goal of creating a “human-like” voice agent and the ethical and legal requirement for transparency.1 The more successful the technology is at mimicking human speech and conversation, the more critical it becomes to explicitly and clearly state that the caller is speaking with a machine. A truly expert implementation will not only build a fluid, low-latency conversation but will also ensure that the AI delivers a clear, early disclosure to the caller, balancing technological aspiration with legal and ethical responsibility.8</p><h3>7.0 Conclusion: The Future is Conversational</h3><p>The integration of Twilio’s global voice network and OpenAI’s GPT Realtime API provides a powerful and comprehensive blueprint for a new generation of AI-powered conversational agents. The solution presented herein — a robust, asynchronous server acting as a WebSocket bridge — is capable of delivering a genuinely low-latency, “human-like” experience for both inbound and outbound calls.1 A detailed analysis of the NewOaks AI platform demonstrates that this architecture is not just a theoretical concept but a scalable, commercial product with diverse applications.1</p><p>The key to a successful deployment lies in moving beyond simple API integration to implement sophisticated server-side logic, particularly for handling interruptions and integrating with external systems via tool use.5 By taking these steps, a company can transform a conversational novelty into a powerful business automation tool that books appointments, qualifies leads, and manages customer interactions with unparalleled efficiency. The foundational platform for this new era of conversational AI has been established, and the Twilio-OpenAI stack provides a clear and actionable path for developers and businesses to build upon.</p><h4>Works cited</h4><ol><li>NewOaks AI — AI Agent Reviews, Features, Use Cases &amp; Alternatives (2025), accessed September 12, 2025, <a href="https://aiagentsdirectory.com/agent/newoaks-ai">https://aiagentsdirectory.com/agent/newoaks-ai</a></li><li>How to use GPT Realtime API for speech and audio with Azure OpenAI in Azure AI Foundry Models — Microsoft Learn, accessed September 12, 2025, <a href="https://learn.microsoft.com/en-us/azure/ai-foundry/openai/realtime-audio-quickstart">https://learn.microsoft.com/en-us/azure/ai-foundry/openai/realtime-audio-quickstart</a></li><li>User Guides Intro | Help Docs — NewOaks AI, accessed September 12, 2025, <a href="https://www.newoaks.ai/guides/user/intro">https://www.newoaks.ai/guides/user/intro</a></li><li>Media Streams Overview | Twilio, accessed September 12, 2025, <a href="https://www.twilio.com/docs/voice/media-streams">https://www.twilio.com/docs/voice/media-streams</a></li><li>twilio-samples/speech-assistant-openai-realtime-api-node — GitHub, accessed September 12, 2025, <a href="https://github.com/twilio-samples/speech-assistant-openai-realtime-api-node">https://github.com/twilio-samples/speech-assistant-openai-realtime-api-node</a></li><li>Advanced AI Solutions &amp; Services | NewOaks — JustCall, accessed September 12, 2025, <a href="https://justcall.io/ai-agent-directory/newoaks-ai/">https://justcall.io/ai-agent-directory/newoaks-ai/</a></li><li>LiveChat vs NewOaks AI, accessed September 12, 2025, <a href="https://www.newoaks.ai/blog/best-livechat-alternative-comparison/">https://www.newoaks.ai/blog/best-livechat-alternative-comparison/</a></li><li>Make Outbound calls with Python, the OpenAI Realtime API, and Twilio Voice, accessed September 12, 2025, <a href="https://www.twilio.com/en-us/blog/outbound-calls-python-openai-realtime-api-voice">https://www.twilio.com/en-us/blog/outbound-calls-python-openai-realtime-api-voice</a></li><li>AI Voice Assistant with Twilio Voice, OpenAI’s Realtime API, and …, accessed September 12, 2025, <a href="https://www.twilio.com/en-us/blog/voice-ai-assistant-openai-realtime-api-python">https://www.twilio.com/en-us/blog/voice-ai-assistant-openai-realtime-api-python</a></li><li>Docs: API Reference, Tutorials, and Integration | Twilio, accessed September 12, 2025, <a href="https://www.twilio.com/docs">https://www.twilio.com/docs</a></li><li>Consume a real-time Media Stream using WebSockets, Python, and Flask | Twilio, accessed September 12, 2025, <a href="https://www.twilio.com/docs/voice/tutorials/consume-real-time-media-stream-using-websockets-python-and-flask">https://www.twilio.com/docs/voice/tutorials/consume-real-time-media-stream-using-websockets-python-and-flask</a></li><li>OpenAI Realtime API Documentation, accessed September 12, 2025, <a href="https://platform.openai.com/docs/guides/realtime">https://platform.openai.com/docs/guides/realtime</a></li><li>Build an AI Voice Assistant with Twilio Voice, OpenAI’s Realtime API, and Node.js, accessed September 12, 2025, <a href="https://www.twilio.com/en-us/blog/voice-ai-assistant-openai-realtime-api-node">https://www.twilio.com/en-us/blog/voice-ai-assistant-openai-realtime-api-node</a></li><li>Pricing plans for websites of all sizes — NewOaks AI, accessed September 12, 2025, <a href="https://www.newoaks.ai/pricing">https://www.newoaks.ai/pricing</a></li><li>Build a RAG-Powered Voice Agent with Twilio Voice, OpenAI, Astra DB, and Node.js | DataStax, accessed September 12, 2025, <a href="https://www.datastax.com/blog/rag-voice-agent-twilio-openai-astra-db-node-js">https://www.datastax.com/blog/rag-voice-agent-twilio-openai-astra-db-node-js</a></li><li>NewOaks AI Integration with FlowMattic, accessed September 12, 2025, <a href="https://flowmattic.com/integration/newoaks-ai/">https://flowmattic.com/integration/newoaks-ai/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4039d85c3bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Build a Conversational AI Voice Agent with GPT Realtime API in 2025]]></title>
            <link>https://medium.com/@luanxr2000/how-to-build-a-conversational-ai-voice-agent-with-gpt-realtime-api-in-2025-fb4243af419e?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb4243af419e</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Fri, 05 Sep 2025 16:04:44 GMT</pubDate>
            <atom:updated>2025-09-05T16:04:44.587Z</atom:updated>
            <content:encoded><![CDATA[<h3>Executive Summary</h3><p>The development of a conversational AI voice agent that leverages Retrieval-Augmented Generation (RAG) and the OpenAI GPT Realtime API is a strategic undertaking that combines advanced linguistic processing with proprietary knowledge systems. The core challenge in this endeavor is the inherent architectural constraint of the GPT Realtime API, which, while revolutionary in its low-latency, speech-to-speech capabilities, does not natively support direct data retrieval from external knowledge bases. This necessitates a hybrid architectural model.</p><p>The optimal solution involves utilizing a specialized platform like NewOaks.ai as a front-end interface and a pre-built conversational framework, while concurrently developing a dedicated, custom backend for RAG functionality. This backend “Orchestration Layer” is responsible for securely retrieving relevant information from a curated knowledge base and injecting it into the conversation context for the GPT Realtime API. This approach minimizes redundant development work on the front-end user experience and core conversational management while providing complete control over the critical data retrieval, security, and scalability components.</p><h3><strong>Key Findings:</strong></h3><ul><li><strong>RAG is Foundational for Factual Accuracy:</strong> RAG is an architectural necessity for any enterprise-grade voice agent. It grounds generative models in verifiable, up-to-date, and domain-specific information, thereby mitigating the risk of factual inaccuracies and hallucinations that plague models reliant on static training data.</li><li><strong>GPT Realtime API’s Latency is a Game-Changer:</strong> The API’s direct speech-to-speech processing pipeline eliminates the multi-step conversion chain of traditional voice assistants, resulting in significantly reduced latency and the preservation of crucial speech nuances like intonation and emotion. This is the single most important technical breakthrough for enabling a truly natural, human-like conversational experience.</li><li><strong>NewOaks.ai Provides a Strong Starting Point:</strong> NewOaks.ai offers a pre-packaged AI Voice Web Call Widget and a proven RAG system for text-based chatbots, which can be a valuable asset for rapid deployment. However, the platform’s internal RAG system cannot be seamlessly integrated with the GPT Realtime API without a custom, external retrieval layer.</li><li><strong>A Custom Orchestration Layer is Required:</strong> The GPT Realtime API is a stateless endpoint that relies on the client to provide all conversational context. To implement a RAG system, an application must perform the data retrieval externally and pass the retrieved information to the API. This external retrieval component is critical for maintaining data isolation and security, especially when handling user-specific information.</li></ul><h3><strong>Strategic Recommendation:</strong></h3><p>The recommended path forward is to adopt a hybrid architecture. Leverage NewOaks.ai’s established front-end voice widget and integrations, while custom-building a dedicated, scalable backend to manage the RAG pipeline. This strategy allows the development team to focus on the most complex and differentiating component — the knowledge management and retrieval system — while benefiting from the proven conversational and integration features of a third-party platform. This hybrid model offers the fastest route to a production-ready, factually-grounded, and low-latency AI voice agent.</p><h3>NewOaks.ai: An Analysis of Platform Capabilities</h3><h3>NewOaks.ai’s Role in the Voice AI Ecosystem</h3><p>NewOaks.ai is a specialized AI agent and chatbot builder designed to automate customer engagement, lead generation, and appointment scheduling. The platform provides an intuitive, web-based interface that enables businesses to create conversational AI solutions without extensive technical expertise. Its core offerings include an AI Voice Web Call Widget, omnichannel communication support, and bi-directional integrations with platforms like Twilio, Google Voice, and CRM systems such as Salesforce and HubSpot.</p><p>A key feature of NewOaks.ai is its existing RAG system, which is utilized for its text-based chatbots. The platform allows users to train their bots by uploading various document types, including TXT, PDF, and DOCX files, as well as providing website URLs and custom Q&amp;A pairs. This pre-built RAG system is a core part of its value proposition, enabling chatbots to provide real-time, accurate, and factually grounded responses. One user review explicitly notes that the NewOaks AI RAG system addresses the problem of “efficiently managing and storing data for real-time, accurate chatbot responses”.</p><p>For voice-based applications, NewOaks.ai offers a dedicated “AI Voice Web Call Widget” that visitors can embed on their website to initiate instant, natural voice conversations directly through their browser. This widget is explicitly stated to be “Powered by ChatGPT’s real-time audio API,” establishing a direct link between the NewOaks.ai platform and the core technology required for low-latency voice AI. This suggests that NewOaks.ai has already handled the complex front-end development of integrating with the GPT Realtime API, which is a significant time-saver for a business seeking a rapid deployment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0j-QH_60Kk19_Od-Mj5Cmw.png" /><figcaption>AI Voice Call Widget by NewOaks AI</figcaption></figure><p>The platform’s versatility is further enhanced by its seamless integrations with external services through APIs and Webhooks. These integrations facilitate real-time data synchronization with CRM systems, enable automated SMS messaging via Twilio, and support conversational appointment scheduling with services like Google Calendar. This capability to act as a central hub for various communication channels and data sources makes NewOaks.ai a powerful tool for streamlining customer engagement workflows.</p><h3>Architectural Blueprint: Integrating RAG, GPT Realtime API, and NewOaks.ai</h3><h3>The Foundational Challenge: GPT Realtime API and External Data Retrieval</h3><p>A central misconception in integrating RAG with the GPT Realtime API is that the API can natively connect to and retrieve information from a company’s private knowledge base. The available documentation from Microsoft and OpenAI makes it explicitly clear that this is not the case. The GPT Realtime API is a conversational endpoint designed for low-latency inference; it cannot, by itself, fetch or access files from a blob storage or a vector database.</p><p>This architectural constraint means that the retrieval of user-specific data must be performed by an external application or service. This external component acts as an intermediary, or an “Orchestration Layer,” between the front-end user interface and the Realtime API. This layer is an essential part of the system for several critical reasons beyond just technical necessity. It is the component responsible for maintaining security and proper data isolation. It securely connects to the knowledge base, fetches the required user data, and handles authentication and authorization. By performing these tasks on a secure backend, the system ensures that a single user’s query cannot inadvertently retrieve or access information from another user’s files, thereby preventing data mixing and maintaining privacy.</p><p>Therefore, any RAG implementation with the GPT Realtime API — whether with a custom front-end or a platform like NewOaks.ai — must include a robust and secure backend service to manage the retrieval process. The front-end’s role is simply to capture the user’s audio and send it to this backend, and the backend’s role is to handle the complex logic of retrieval and prompt augmentation before passing the conversation to the Realtime API.</p><h3>Blueprint for a Custom Retrieval Layer</h3><p>The architectural blueprint for a fully functional, RAG-powered voice agent is a multi-component system that intelligently orchestrates the flow of information between the user, the knowledge base, and the generative model. The following table provides a breakdown of the key components and their roles in this hybrid architecture.</p><p>The workflow of the system is a continuous, multi-step process:</p><ol><li><strong>User Initiation:</strong> A user clicks on the NewOaks.ai AI Voice Web Call Widget on a website, initiating a WebRTC or WebSocket connection with the custom Orchestration Layer.</li><li><strong>Audio Streaming:</strong> The user’s audio is streamed in real time to the Orchestration Layer.</li><li><strong>Realtime Transcription &amp; Retrieval:</strong> The Orchestration Layer forwards the audio to the GPT Realtime API, which returns real-time transcript deltas. The Orchestration Layer uses these transcripts to perform a semantic search against the vector database. A critical architectural choice is whether to initiate this search on a full utterance (when</li><li>semantic_vad detects a pause) or proactively using the streaming text. The latter option can reduce overall latency but carries a risk of retrieving irrelevant information if the user&#39;s intent changes mid-sentence.</li><li><strong>Prompt Augmentation:</strong> Once the relevant information is retrieved from the vector database, the Orchestration Layer constructs an augmented prompt. This new prompt combines the user’s query with the factual context from the knowledge base.</li><li><strong>Grounded Generation:</strong> The augmented prompt is sent to the GPT Realtime API. The model, now equipped with external, verifiable facts, generates a grounded response.</li><li><strong>Realtime Audio Synthesis:</strong> The Realtime API synthesizes the response and streams the audio back to the Orchestration Layer.</li><li><strong>Final Audio Delivery:</strong> The Orchestration Layer streams the synthesized audio back to the NewOaks.ai widget, which plays the response to the user.</li></ol><p>This blueprint addresses the foundational challenge directly. By treating the GPT Realtime API as a powerful but stateless voice engine, and building a secure, custom-built RAG backend, the solution achieves both the low-latency, natural conversation of the API and the factual accuracy and security of a proprietary RAG system. The NewOaks.ai widget acts as the essential front-end interface, saving significant development time and effort.</p><h3>Implementation, Optimization, and Best Practices</h3><h3>Building the RAG Pipeline</h3><p>The effectiveness of any RAG-powered system is directly proportional to the quality of its underlying knowledge base and the efficiency of its retrieval mechanism. To build a robust pipeline, several best practices must be followed.</p><p>First, the knowledge base must be curated with comprehensive and up-to-date content that is relevant to user inquiries. This is a continuous process, as information can become stale over time. Regular updates are essential to ensure the agent provides the latest information. The data should also be structured and indexed effectively to facilitate quick and accurate retrieval. Using metadata and tagging is a valuable technique for enhancing search efficiency, allowing for more precise filtering and targeting of information. A robust chunking strategy is also vital; breaking down large documents into smaller, meaningful chunks ensures that the retrieved context is concise and highly relevant, preventing the LLM from being overwhelmed by extraneous information.</p><p>Second, the retrieval mechanism must be optimized. While a basic semantic search is a starting point, a more sophisticated approach is required for production-grade performance. Hybrid search, which combines the precision of keyword search with the conceptual understanding of semantic search, is highly recommended. Implementing a relevancy re-ranker further refines the search results, ensuring that the most relevant documents are returned to the LLM. The choice of embedding models is also paramount. The model should be carefully selected to accurately represent the content domain of the knowledge base, as a mismatch can lead to poor retrieval performance.</p><h3>Designing for a Natural, Low-Latency Conversation</h3><p>The GPT Realtime API provides the technical foundation for a natural conversation, but achieving this requires careful configuration and a focus on user experience.</p><p>The most critical configuration is the turn_detection setting, which dictates how the system determines when a user has finished speaking. The API offers two main modes:</p><p>server_vad (Voice Activity Detection), which chunks audio based on periods of silence, and semantic_vad, which uses the model&#39;s understanding of the words to determine the end of an utterance. The semantic_vad mode is the superior choice for a natural conversation, as it is less likely to interrupt the user mid-sentence and allows for a more fluid dialogue flow. Additionally, the system must be designed to handle user interruptions gracefully. A truly natural conversation allows for one party to interrupt the other, and the voice agent must be configured to stop its response and listen for new input when a user begins to speak.</p><p>A strategic comparison between the new technology and traditional methods highlights its profound impact on the user experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/804/1*qAznaTM5iN3rjp18S3pzWA.png" /></figure><p>This table demonstrates that the GPT Realtime API provides a direct path to the qualities that define a positive user experience. The ability to preserve speech nuances and respond to non-verbal cues means the agent can adapt its tone and pacing, making it more engaging and empathetic. The success of the project is not just a technical measure of speed or accuracy, but a perceptual measure of “naturalness.”</p><h3>4.3. Cost and Scalability Management</h3><p>Managing costs for a real-time voice agent is a critical consideration. Unlike text-based APIs, the GPT Realtime API’s pricing is based on audio input and output tokens, which have different rates. The platform offers a 20% price reduction compared to previous models, with rates of $32 per 1M audio input tokens and $64 per 1M audio output tokens. This pricing structure necessitates specific cost optimization strategies.</p><p>To manage costs effectively, the following measures are recommended:</p><ul><li><strong>Intelligent Token Limits:</strong> The API includes features for fine-grained conversation context control. By setting reasonable limits on the context length, a business can reduce costs for long or rambling conversations.</li><li><strong>Multi-turn Truncation:</strong> For extended sessions, the ability to truncate multiple conversation turns at once can significantly reduce token usage and, consequently, costs.</li><li><strong>Model Selection:</strong> For simpler, high-volume use cases, leveraging the more cost-effective gpt-4o-mini-realtime-preview model, which is a less expensive version of the core model, can yield substantial savings. For complicated use cases like lead capture and appointment booking, Gpt-Realtime mode is preferred.</li><li><strong>Prompt Caching:</strong> For frequently asked questions, caching prompts and responses can reduce the need for repeated API calls, thereby lowering costs and response times.</li></ul><p>The architectural blueprint with a custom Orchestration Layer is inherently more scalable than a simple client-side solution. By deploying the RAG backend on a scalable cloud infrastructure (e.g., Azure or AWS), the system can handle a high volume of concurrent voice sessions without performance degradation. The backend is responsible for securely managing user authentication and data retrieval at scale, ensuring that the system can grow with business needs.</p><h3>Strategic Recommendations and Conclusion</h3><h3>The Value Proposition: Benefits and ROI</h3><p>The implementation of a RAG-powered, low-latency AI voice agent provides a clear and compelling value proposition for businesses. The system’s ability to provide instant, accurate, and personalized support 24/7 represents a significant improvement in customer experience and a major reduction in operational overhead. Businesses using AI in customer service have reported a 37% drop in first response times and a 55% decrease in wait times, highlighting the efficiency of these systems.</p><p>The financial return on investment is substantial. By automating a high percentage of customer inquiries, businesses can cut customer service costs by up to 50%. The agent’s ability to handle both inbound calls and outbound lead qualification automates high-volume tasks without compromising quality or requiring additional human resources. Furthermore, the system can capture valuable conversational data, enabling sentiment analysis and feedback that can drive improvements in product development and customer satisfaction. This transformation turns customer support from a cost center into a source of strategic insights.</p><h3>A Final Assessment: Build vs. Buy</h3><p>The initial query, “how to use RAG to build a conversational AI Voice agent with NewOaks.ai integration of GPT Realtime API,” presents what appears to be a straightforward integration task. However, a deeper technical analysis reveals it to be a strategic “build vs. buy” decision. A purely “buy” solution using only NewOaks.ai is not feasible for a custom, fact-grounded agent because the GPT Realtime API’s technical architecture necessitates a custom, external retrieval layer for RAG.</p><p>Therefore, the most robust and strategic path forward is a hybrid model. The “buy” component, NewOaks.ai, provides the valuable front-end voice widget and a user-friendly platform for managing workflows and integrations. The “build” component, the custom RAG orchestration layer, is where the core value and technical complexity of the project reside. This component is crucial for handling the secure, real-time data retrieval that provides the agent with its unique, fact-based intelligence.</p><h3>Final Recommendations</h3><p>Based on this comprehensive analysis, the following recommendations provide a clear path forward for the implementation of a production-grade, RAG-powered voice agent:</p><ol><li><strong>Adopt a Hybrid Architecture:</strong> The most effective and robust solution is a hybrid model that leverages NewOaks.ai’s pre-built AI Voice Web Call Widget and its conversational framework while developing a dedicated, scalable backend for the RAG pipeline. This minimizes front-end development time and ensures complete control over the most critical aspects of the agent’s intelligence and security.</li><li><strong>Prioritize Knowledge Base Curation:</strong> The success of the RAG system hinges on the quality of its data. Before writing any code, invest significant resources in curating a high-quality, well-structured, and continuously updated knowledge base. This will be the primary source of the agent’s accuracy and value.</li><li><strong>Plan for Ongoing “RAG Ops”:</strong> Recognize that this is a product lifecycle, not a one-time project. Implement a continuous evaluation loop to monitor metrics like groundedness, latency, and user satisfaction. The system will require ongoing maintenance, data updates, and algorithm refinements to ensure its continued performance and relevance.</li></ol><p>The convergence of RAG and the GPT Realtime API presents an unprecedented opportunity to create truly intelligent, natural, and effective conversational voice agents. By understanding the core technical constraints and adopting a strategic hybrid architecture, a business can leverage these technologies to create a solution that is both highly effective and a significant competitive differentiator.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb4243af419e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Clone Nano Banana Website with AI in 2025]]></title>
            <link>https://medium.com/@luanxr2000/how-to-clone-nano-banana-website-with-ai-in-2025-586f49c4e1b2?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/586f49c4e1b2</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Sun, 31 Aug 2025 15:17:06 GMT</pubDate>
            <atom:updated>2025-08-31T15:17:06.612Z</atom:updated>
            <content:encoded><![CDATA[<p>Many people are puzzled after seeing <a href="http://nanobanan.ai/">Nanobanan.ai</a>:</p><p>How can such an amazing website be built? No problem, this article is a hands-on tutorial! I’ll take you from zero to recreate this viral website with no prior experience, helping you not only understand the logic behind it but also build your own AI image generation site.</p><p>Today, our goals are clear:</p><ul><li>Integrate Gemini 2.5 Flash image Preview model (text-to-image &amp; image-to-image)</li><li>Build a complete Landing Page</li><li>Make it ready for real deployment</li></ul><p>This article will completely replicate the deployment process — just follow along and you’ll get it working.</p><h3>Prompt Construction</h3><p>We now need to go to <a href="http://easysite.ai/">Easysite.ai</a> (or any other AI programming tools like Cursor, Claude Code, Bolt, Lovable, etc. — the principle is the same) to send cloning instructions. Try to prepare the prompt content as comprehensively as possible, giving AI everything needed for design and development in one go.</p><p>Here we need to think about what content to include in our prompts. Prompts aren’t about rote memorization, but about designing and assembling them based on our needs. Our requirements here are: Clone Target: We need to replicate as close to 100% as possible, so the content we provide should ideally be in this order of effectiveness: Code &gt; Screenshots</p><p>To simplify the operation, we’ll use the “clone by URL” approach to provide the AI image generation model, replicating Nano Banana through code parsing + screenshots.</p><p>Find APIs for online model providers and provide input examples and output examples, so the AI knows the format and calling requirements.</p><h3>Technical Solution:</h3><p>For better results and easier maintenance later, we commonly use React/Vue + Node Server. For other technical solutions, you can use Chat Mode to ask AI.</p><p>This section is explaining the strategic approach to creating effective prompts for AI development tools:</p><p>Priority of reference materials: Actual code is better than screenshots for replication Method: Using URL-based cloning with code parsing and visual references</p><h3><strong>API Integration:</strong></h3><p>Need to find and document API endpoints with clear input/output examples Tech Stack: Recommending modern web development stack (React/Vue frontend with Node.js backend)</p><h3><strong>Model Search: </strong>Gemini 2.5 Flash Image Preview</h3><p>Image Editor is an image-to-image model that requires uploading an image;</p><p>The base model is text-to-image.</p><p><strong>Get input examples and output examples:</strong></p><p>Finally, generate your own API KEY in your personal account, and you can combine it into the model context we need:</p><ul><li>Get Your Own API KEY: <a href="https://aistudio.google.com/">https://aistudio.google.com/</a></li><li>Image-to-image model (example input + example output)</li></ul><p><a href="https://ai.google.dev/gemini-api/docs/image-generation#rest">Image generation with Gemini (aka Nano Banana) | Gemini API | Google AI for Developers</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/894/1*ZCpWAHolXPgO1MCYT_jXKg.png" /></figure><ul><li>Text-to-image model (example input + example output)</li></ul><p><a href="https://ai.google.dev/gemini-api/docs/image-generation#rest">Image generation with Gemini (aka Nano Banana) | Gemini API | Google AI for Developers</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/925/1*p_7Opcn9l9VoDF_jNGfskw.jpeg" /></figure><p>This section is outlining the technical components needed for the AI image generation integration:</p><p><strong>Model Identification</strong>:</p><p>Specifying the exact Gemini 2.5 Flash model (Gemini 2.5 Flash Image Preview)</p><ul><li>Edit version: Requires image upload for image-to-image generation</li><li>Base version: Text-to-image generation</li></ul><p><strong>Documentation Requirements</strong>:</p><p>Need to gather:</p><ul><li>API authentication (API key)</li><li>Working examples of both model types with their input/output formats</li></ul><h3>Step by Step Clone Nano Banana on EasySite.AI:</h3><ol><li><a href="http://1.open/">Open</a> <a href="http://easysite.ai/">Easysite.ai</a></li></ol><p>2. Select Clone Tab: Enter <a href="https://nanobanana.ai/">https://nanobanana.ai</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kY0Kmn8L5qDQ0BJpZT2aFQ.png" /></figure><p>3. Select all sub pages that you want to clone</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jiXNbuBzBV_JEaGMiO0Jow.png" /></figure><p>4. Use following prompt:</p><ol><li>Remove Batch Processing feature</li></ol><p><strong>2. Images can be uploaded up to a maximum of 9 pictures</strong></p><p><strong>3. Pay attention to the dynamic interactive parts on the website, such as proper tab implementation and maintaining consistent active states</strong></p><p><strong>4. Unified expand/collapse interactions for the FAQ section</strong></p><p><strong>5. Unified image upload component interactions, especially for multiple image uploads and deletion of uploaded images</strong></p><p><strong>6. The second screenshot shows the style after switching to the text-to-image tab</strong></p><p><strong>7. Technology:</strong></p><p>Tech stack uses React + Tailwind CSS + Node Server</p><p><strong>8. Image Generation Models:</strong></p><p>gemini-2.5-flash-image-preview</p><p>9. Use this GEMINI API KEY: &lt;YOUR API KEY&gt;;</p><p>10. implement the API key in node-server;</p><p>use following API for text to image feature:</p><p>curl -s -X POST”<a href="https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent">https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent</a>&quot; \-H “x-goog-api-key: $GEMINI_API_KEY” \-H “Content-Type: application/json” \-d ‘{“contents”: [{“parts”: [{“text”: “Create a picture of a nano banana dish in a fancy restaurant with a Gemini theme”}]}]}’ \| grep -o ‘“data”: “[^”]*”’ \| cut -d’”’ -f4 \| base64 — decode &gt; gemini-native-image.png ;</p><p>Use following Rest API for image to image feature:</p><p>IMG_PATH=/path/to/cat_image.jpegif [[ “$(base64 — version 2&gt;&amp;1)” = <em>“FreeBSD”</em> ]]; thenB64FLAGS=” — input”elseB64FLAGS=”-w0&quot;fiIMG_BASE64=$(base64 “$B64FLAGS” “$IMG_PATH” 2&gt;&amp;1)curl -X POST \&quot;<a href="https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent">https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent</a>&quot; \-H “x-goog-api-key: $GEMINI_API_KEY” \-H ‘Content-Type: application/json’ \-d “{\”contents\”: [{\”parts\”:[{\”text\”: \”’Create a picture of my cat eating a nano-banana in a fancy restaurant under the Gemini constellation\”},{\”inline_data\”: {\”mime_type\”:\”image/jpeg\”,\”data\”: \”$IMG_BASE64\”}}]}]}” \| grep -o ‘“data”: “[^”]*”’ \| cut -d’”’ -f4 \| base64 — decode &gt; gemini-edited-image.png</p><p>Here is the cloned landing page:</p><p><a href="https://nanobanana-sites.easysite.ai/">https://nanobanana-sites.easysite.ai/</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/943/1*r1jrvbNU_nFak7KN2fbC_w.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=586f49c4e1b2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 5 AI Website Cloning Tools For Non-Developer in 2025]]></title>
            <link>https://medium.com/@luanxr2000/top-5-ai-website-cloning-tools-for-non-developer-in-2025-891b3cfad868?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/891b3cfad868</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 15:10:56 GMT</pubDate>
            <atom:updated>2025-08-21T15:10:56.026Z</atom:updated>
            <content:encoded><![CDATA[<p>By Ray Luan</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fo7bCLojD5iKc-6E.png" /></figure><p>Looking for an easy way to clone any UI or website without learning code? Here are the top 5 clone website ai tools for non-developers in 2025: <strong>EasySite AI, KwikUI, 10Web AI, Lovable.dev, and Bolt.new.</strong> These ai web development tools use artificial intelligence to <a href="https://www.digitalocean.com/resources/articles/ai-tools-web-development">handle design, coding, and content creation</a> for you.</p><p>Simply enter a website link, and these clone website ai solutions do the rest. With AI website cloning tools, you can quickly <a href="https://www.linkedin.com/pulse/impact-ai-web-development-opportunities-challenges-mirza-hadi-baig-xjm4f">clone any UI and build impressive websites in no time</a>. Enjoy fast website creation, effortless editing, and <a href="https://www.webmastered.com/blog/ai-in-web-development-how-agencies-are-adapting/">intelligent layout suggestions</a>. Now, anyone can use clone website ai to make stunning sites and clone any UI within minutes.</p><h3>Key Takeaways</h3><ul><li>AI website cloning tools help people copy websites fast. You do not need to know how to code. These tools copy the look, words, and parts you can click. You can change things easily with drag-and-drop builders. Top tools like EasySite AI, KwikUI, 10Web AI, Lovable.dev, and Bolt.new have special features. They work for different needs and prices. Pick the best tool by looking at how easy it is to use. Think about the price, features, ways to change things, and help you get. Using AI website cloning tools saves you time and money. You can make good websites fast and safe.</li></ul><h3>What Are AI Website Cloning Tools?</h3><p>Have you ever wanted to copy a website without coding? AI website cloning tools help anyone do this, even if you do not know tech stuff. These tools use artificial intelligence to copy websites, layouts, and user interfaces with just a few clicks. You can use features that only developers had before. Now, you can <a href="https://aloa.co/ai/resources/industry-insights/build-ai-websites">make, change, and launch websites easily</a>.</p><h3>How Clone Website AI Works</h3><p>Clone website ai looks at the site you want to copy. It takes all the words, pictures, and design parts. Then it rebuilds them using smart ai web development tools. You just type in a website link or say what you want. The ai website generator checks the real site, copies the layout, and makes things like forms and buttons work the same way. You can check the main features in the table below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/461/1*Ok8fk5n4nJfasNVo0zOD0w.png" /></figure><h3>Why Clone Any UI with AI?</h3><p>Why use ai website cloning tools instead of old ways? The answer is they are fast, correct, and flexible. Clone website ai lets you <a href="https://medium.com/%40learnwithwhiteboard_digest/7-top-ai-website-cloning-tools-for-frontend-developers-f2793c4f900b">copy any UI in minutes</a>. This saves you lots of time. You can make new pages, test ideas, or change your brand quickly. AI website builders help you avoid mistakes and keep your site looking perfect. You also save money because you do not need a developer for every change. With ai web development tools, you can change your copied site anytime, try new looks, and test ideas easily. These ai tools for web development also keep your SEO strong, so your site stays easy to find online.</p><blockquote>Tip: AI website cloning tools use <a href="https://www.businessnewsdaily.com/ai-security-risk">strong privacy and encryption</a> to keep your data safe. You control your information while using artificial intelligence.</blockquote><p>If you want to copy any UI or website fast, ai website cloning tools are the best pick. They let you make, change, and run sites like an expert, even if you are new.</p><h3>Top 5 AI Website Cloning Tools for 2025</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KFSG9vavIUOwMIjnfvCEgQ.jpeg" /></figure><p>Picking the best ai website cloning tools can be hard if you do not code. <a href="https://medium.com/what-is-the-best-ai/best-ai-website-builders-2025-i-tested-20-tools-to-find-the-top-performers-5794c9879d32">I tried more than 20 platforms by making the same website on each</a> one. I checked how simple they are, how well they copy designs, how fast they work, and how much you can change things. I also looked at their prices, support, and how good they are for SEO. <a href="https://blog.getbind.co/2025/07/06/the-5-best-ai-website-builders-in-2025/">Here is a table that shows what I looked for</a>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/1*49GtSaqFHnjaKxtXSBCzmA.png" /></figure><p>Let’s look at the top 5 ai website cloning tools for 2025. Each tool has special features and helps you make a business website fast without coding.</p><h3>EasySite AI</h3><p><a href="http://www.easysite.ai/">EasySite AI </a>is a top ai website generator. You type in a website link, and the ai makes a copy as a full app. You get a drag-and-drop builder, so you can change things by moving them around. <a href="https://www.ezsite.ai/blog/best-lovable-alternative-2025/">EasySite AI works very fast. It can make a new website or app in about one minute</a>. You do not need to know any tech stuff.</p><p>You get more than just a copy of the look. EasySite AI has a database, login, AI chatbots, tools for leads and payments, API links, hosting, custom web addresses, stats, and teamwork tools. This is great for small businesses, new companies, freelancers, and big companies that want something that grows with them.</p><blockquote>Note: EasySite AI has a free trial plan and cheap paid plans with Pay Per Usage feature. There are no secret costs, and starter plans cost less than most others.</blockquote><p>Here is a table with what you get:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*9D7LTM1nUg8F9PLyQrd5fw.png" /></figure><p>If you want to clone a site with one click and use a builder for everything, EasySite AI is a great choice.</p><h3>KwikUI</h3><p>KwikUI works in a different way. You upload a screenshot or picture of any UI you like. <a href="https://blog.kwikui.com/what-is-kwikui-generate-ui-code-prompts-from-screenshots-explained-2d077bbfe5dd">KwikUI’s ai turns that picture into a smart prompt for other ai website builders</a> like Lovable, Cursor, or Replit. This lets you turn any design into code fast without using an IDE.</p><ul><li><a href="https://www.kwikui.com/">You do not need to know how to code</a>.</li><li>Upload your screenshot, and KwikUI gives you a prompt to use.</li><li>The result works with many ai website generator platforms.</li><li>You get checks for colors, spacing, and fonts.</li><li>KwikUI works in your browser, so there is nothing to install.</li></ul><p><a href="https://blog.kwikui.com/difference-between-low-code-vs-ai-code-generation-which-builds-an-mvp-faster-4b960f39cd74">Many people say KwikUI helps them build UIs much faster</a>. It is good for designers, product managers, and anyone who does not want to code by hand. <a href="https://medium.com/%40learnwithwhiteboard_digest/12-top-ai-tools-for-mvp-development-to-move-fast-c7bc7415d2ef">The free plan lets you try three screenshots, and paid plans start at $9.99 a month</a>.</p><blockquote>Tip: KwikUI is great if you want to turn ideas into code quickly and use different ai website builders for your projects.</blockquote><h3>10Web AI</h3><p><a href="https://siteefy.com/ai-tools/10web/">10Web AI is a popular ai website builder that lets you copy sites or make new ones</a>. It uses WordPress and Google Cloud for hosting. You get a drag-and-drop editor, so you can change your site easily. 10Web AI also makes text and pictures for you.</p><p>All plans have free SSL, moving help, 24/7 support, and a 7-day free trial. <a href="https://calanbreckon.com/10web/">People like how 10Web AI copies the look and layout of a site</a>, <a href="https://msbdigital.com/10web-ai-review/">saving lots of time</a>. You can change everything with the drag-and-drop builder, so it is easy for anyone.</p><blockquote>Note: <a href="https://appsumo.com/products/10web-ai-website-builder/reviews/it-takes-a-lot-for-me-to-leave-a-negativ-101538/">Some people say you might need to fix links or add pages</a>, but the ai website generator does most of the work.</blockquote><h3>Lovable.dev</h3><p><a href="https://www.sidetool.co/post/how-does-lovable-ai-compare-to-no-code-platforms-like-v0-and-windsurf/">Lovable.dev is a top ai website cloning tool for people who do not code</a>. <a href="https://momen.app/blogs/lovable-dev-pros-cons-ideal-use-cases-for-no-code-apps/">You chat with the ai, and it builds both the front and back parts for you</a>. Lovable.dev works with databases, login, and lets you export React code for more control.</p><ul><li>You get a chat-based, no-code builder.</li><li>Lovable.dev handles hosting, launching, and teamwork in real time.</li><li>You can connect with tools like Supabase, Resend, Vercel, and Netlify.</li><li>It is good for MVPs, test sites, marketing sites, dashboards, and tools for inside your company.</li></ul><p>Plans start at about $20 a month, so most people can afford it. If you want to build a full app or business site with no coding, Lovable.dev is a strong pick.</p><h3>Bolt.new</h3><p><a href="https://uibakery.io/blog/bolt-vs-lovable-vs-v0">Bolt.new is an ai web builder made for speed and teamwork</a>. It works in your browser and lets you build full apps with help from ai. You get teamwork in real time, so you and your team can work together like Google Docs. Bolt.new works with databases, one-click cloud launch, and uses smart AI models like GPT-4 and Claude.</p><p>Bolt.new is great when you need to copy a website fast. <a href="https://www.nocode.mba/articles/bolt-ai-coding">You can clone a homepage like Spotify or Airbnb in minutes</a>, with animations and details. The ai website generator does the design and layout, so you can focus on making it your own. The free plan lets you try basic things, and paid plans give you more tokens and faster work.</p><blockquote>Tip: Bolt.new is best for solo builders, freelancers, and teams who want to make and launch sites quickly with ai website builders.</blockquote><h3>EasySite Overview</h3><h3>Features of This AI Website Development Tool</h3><p>EasySite is an ai website development tool that helps you build modern sites or apps. You can clone any website into a copy you can edit. Just use a link to get started. It has a database and CRM built in, so you can keep track of products or customers without coding. The AI agent helps set up things like appointments, payments, and chatbots. You can use agent mode to do tasks for you or connect with other tools. The drag-and-drop website builder lets you move things and see changes right away.</p><ul><li>Fullstack app cloning gives you quick results</li><li><a href="https://lifetimo.com/deal/ezsite-lifetime-deal/">Built-in database and CRM</a> help manage your data</li><li>AI agent and agent mode make smart tasks easy</li><li>Drag-and-drop website builder lets you edit fast</li><li>Visual editing helps you design your site perfectly</li></ul><h3>How to Clone Any Website with EasySite AI</h3><p>Cloning a website with EasySite AI is easy. Type in the URL of the site you want to copy. The ai scans the site and makes a copy you can edit in about a minute. You do not need to know code. The builder shows your new site clearly, so you can start editing fast. You can add forms, change layouts, or update images with a few clicks. The drag-and-drop website builder lets you move parts, change colors, and fix text easily.</p><h3>Visual Editing and Customization</h3><p>You can control how your site looks and feels. The visual editor lets you drag, drop, and resize anything on the page. If you want to change a button or add a section, just move it. The ai gives you layout and color ideas that fit your brand. You can use the builder to add animations or interactive things. Everything updates right away, so you see your changes instantly.</p><h3>Pricing and Plans</h3><p>EasySite has a free plan if you want to try the ai website development tool. Paid plans start at a low monthly price and give you more features like better CRM, more storage, and faster support. You can choose a plan that fits your needs, whether you have a small business or many sites.</p><h3>Best Use Cases</h3><p>EasySite is good for many people. Marketing agencies can clone client sites and launch campaigns quickly. Product managers use the ai to test ideas without waiting for developers. Developers and UI/UX designers like the drag-and-drop website builder for fast prototypes. Website builders and SMBs use EasySite to manage products, leads, and payments with the built-in database and CRM. You can even turn your site into a mobile app with one click.</p><h3>How to Choose an AI Web Development Tool</h3><h3>Key Factors for Non-Developers</h3><p>Choosing the right ai web development tools can feel overwhelming, especially if you’re new to no-code development. You want a builder that makes things simple and fun. Here are some key things you should look for:</p><ul><li><strong>Ease of use:</strong> Pick a <a href="https://betanews.com/best-ai-website-builders/">drag-and-drop website builder</a> with a user-friendly interface. You should feel comfortable moving things around and editing your site without help.</li><li><a href="https://www.digitalocean.com/resources/articles/ai-tools-web-development"><strong>Cost</strong></a><strong>:</strong> Check both the upfront price and any monthly fees. Make sure the tool fits your budget and gives you good value.</li><li><strong>Features:</strong> Look for tools that offer lots of options, like a text-to-ui generator, natural language to website conversion, and built-in SEO tools. Some platforms even let you use a text-to-site generator for faster results.</li><li><strong>Customization:</strong> You want to change colors, layouts, and content easily. The best no-code ai builder lets you make your website look unique.</li><li><strong>Support:</strong> Good customer support matters. Choose an ai web builder with tutorials, live chat, or a helpful community.</li><li><strong>Scalability:</strong> Think about how your site might grow. Some no-code platforms work better for small projects, while others handle bigger sites.</li></ul><blockquote>Tip: <a href="https://allthingsopen.org/articles/2025-guide-ai-no-code-developer-led-software">Match your choice to your goals</a>. If you need a quick launch, go for ai tools for web development that focus on speed and ease.</blockquote><h3>Tips for Cloning Any UI Successfully</h3><p>You can get great results from no-code ai tools if you follow a few simple steps. Here’s how to make your cloned website shine:</p><ol><li>Start with a clear idea of what you want. Use a screenshot or a link to the site you like.</li><li>Pick a drag-and-drop website builder that guides you step by step. This makes the process smooth.</li><li>Use the built-in editing tools to tweak layouts, colors, and text. Real-time previews help you see changes right away.</li><li>Make sure your site works well on phones and tablets. Many ai web builders check this for you.</li><li><a href="https://www.upwork.com/resources/ai-ethical-considerations">Keep your content original</a>. Even if you clone a design, add your own words and images.</li><li>Stay aware of <a href="https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/articles/generative-ai-legal-issues.html">legal and privacy rules</a>. Don’t copy protected content or share sensitive info.</li></ol><h3>FAQ</h3><h3>How do I start cloning a website with these tools?</h3><p>You just pick your favorite tool, enter the link of the site you want, and follow the steps. Most tools show you what to do next. You do not need to know any code.</p><h3>Can I change the design after cloning a site?</h3><p>Yes! You can move things around, change colors, and add your own pictures or text. Most tools have a drag-and-drop editor. You get to make the site look just how you want.</p><h3>Is it safe to use AI website cloning tools?</h3><p>Most tools use strong security to keep your data safe. You control what you share. Always check the privacy settings before you start.</p><h3>Do I need to pay to use these tools?</h3><p>Many tools offer a free plan. You can try them out before you pay. If you want more features, you can choose a paid plan that fits your needs.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=891b3cfad868" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GPT-5 First Impressions: Programming Excellence, Writing Disappointment]]></title>
            <link>https://medium.com/@luanxr2000/gpt-5-first-impressions-programming-excellence-writing-disappointment-588f49e2a589?source=rss-1b2cdc7740ea------2</link>
            <guid isPermaLink="false">https://medium.com/p/588f49e2a589</guid>
            <dc:creator><![CDATA[RayLuan@ EZsite.AI]]></dc:creator>
            <pubDate>Fri, 08 Aug 2025 10:15:40 GMT</pubDate>
            <atom:updated>2025-08-08T10:15:40.934Z</atom:updated>
            <content:encoded><![CDATA[<p>At 1 AM, under the spotlight of global attention, OpenAI’s livestream officially began. GPT-5 had finally arrived. AI has moved so fast that in just 2.5 years, it feels like 10 years have passed. I have to make a decision that whether to integrate GPT-5 into EasySite.AI as a backup for Claude 4.</p><p>On March 15, 2023, GPT-4 was released. In that wild era, everyone was left speechless with shock. Back then, it was the first multimodal model. Back then, everyone thought GPT-5 would come out in the second half of 2023. Back then, everyone held endless aspirations for the upper limits of large language models. As it turned out, we waited 2.5 years.</p><p>After two and a half years of development, GPT-5 finally made its debut. <strong>GPT-5 is a unified system that includes a fast intelligent model for handling most problems (gpt-5-main) and a deep reasoning model designed for high-difficulty problems (gpt-5-thinking)</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZaJ56YxhhCXNrnTJ" /></figure><p>The system uses a real-time router that dynamically selects which model to use based on conversation type, complexity, and user intent. For example, if you say “think carefully about this” in your prompt, it will invoke gpt-5-thinking for deeper consideration.</p><p>This router will continuously train based on our subsequent usage patterns, including cases where users switch models, preference rates for responses, and accuracy measurements, improving over time. The system also includes mini versions for handling excess requests (gpt-5-main-mini and gpt-5-thinking-mini), as well as a smaller, faster nano version designed for developers (gpt-5-thinking-nano). Additionally, there’s a parallel computing version available to Pro members, called gpt-5-thinking-pro. This large system containing so many models is collectively called GPT-5, the direct successor to the previous generation’s GPT-4o and OpenAI o3.</p><p>Here’s a model correspondence table.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*fIRW7wcXZXYn0OaXiRRDyQ.png" /></figure><p>In terms of performance, one of GPT-5’s most significant improvements is the dramatic reduction in factual hallucinations. gpt-5-main produces 44% fewer responses containing at least one major factual error compared to GPT-4o, while gpt-5-thinking produces 78% fewer than OpenAI o3. In more specialized LongFact and FactScore benchmark tests, whether web browsing is enabled or not, GPT-5 series models have significantly lower hallucination rates than their predecessors, with gpt-5-thinking producing five times fewer factual errors than OpenAI o3 in both settings.</p><p>GPT-5 has also made good progress in addressing model sycophancy. Compared to GPT-4o, GPT-5 is less overly accommodating, uses fewer unnecessary emojis, and is more nuanced and thoughtful in subsequent communications. When you chat with it, it feels less like talking to an AI and more like conversing with a friend with PhD-level intelligence. This is very similar to the personalized prompt I give ChatGPT — what I hate most is its sycophantic behavior, so I wrote a section myself to limit its fawning behavior toward me.</p><p>Now through specialized training, gpt-5-main performs nearly three times better than the latest GPT-4o in evaluations. Preliminary online A/B test data shows that compared to GPT-4o, gpt-5-main’s sycophantic behavior occurrence rate dropped by 69% among free users and 75% among paid users.</p><p>They also launched four brand new personality settings, so you don’t need to write lengthy prompts anymore — just change the presets directly. The four are: cynic, robot, listener, and nerd.</p><p>Now let’s look at the benchmark scores. In math competitions, AIME 0225, GPT-4 Pro+Python scored full marks. We need new, more difficult evaluation sets. In real-world programming capabilities, new highs. In humanity’s final knowledge tests, it surpassed ChatGPT Agent, reaching new highs. Multimodal capabilities also hit new highs — basically everything reached new highs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VPAo8_TuuoFSbHDr" /><figcaption>AIME 2025</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FX7VRG4cULpzOlMT" /></figure><p>These scores are stronger, but not dramatically so. Additionally, let me add some complaints — this perfectly demonstrates OpenAI’s slapdash nature. While the charts in the blog are correct, during the launch event, the scores were randomly drawn. For example, 52.8 being greater than 69.1 equals 30.8. Or 50 being less than 47.4. It’s really too slapdash.</p><p>Netizens also spoke up. <strong>Anyway, ultimately GPT-5 dominated the leaderboards across all aspects. The latest large model blind test arena rankings came out, and GPT-5 is comprehensively first place</strong>.<br>Not only is it stronger, it’s also more energy-efficient. In capabilities like visual reasoning, agent programming, and graduate-level scientific problem solving, it performs better than OpenAI o3 while using 50–80% fewer output tokens. Basically, it’s comprehensively stronger across the board.</p><p>But there are no new functions or features. 20 minutes after the release, on Polymarket’s prediction “Which company will have the best AI model by the end of August?”, OpenAI and Google did a synchronized dive.<br>OpenAI said that the entire GPT-5 has made stronger progress than before in writing and programming. <strong>For developers building with GPT-5, pricing is as follows: $1.25 per million tokens (with a 90% caching discount, which is a huge advantage for long context queries). Output: $10 per million tokens.</strong></p><p>After the launch event ended, and after waiting another hour, my friends gradually got access to GPT-5. As a loyal $200 Pro subscriber, I had to wait until 4 AM. When my friends got in and sent me screenshots, my world collapsed. Fine, o3 and 4o are gone, but how did you also kill my GPT-4.5???</p><p>First, in writing and emotional intelligence capabilities, I personally feel it’s still not as good as GPT-4.5… Because I’ve been writing for years and have long used AI to assist with content creation, I consider myself quite sensitive to many subtle details and tones, and GPT-5 still has gaps in this area.<br>For example, a prompt: “If John Doe was scammed by a pretentious and ridiculously expensive coffee shop, what kind of article would he write to complain? Write a short piece under 1000 words.”<br>This is GPT-5’s version. Clumsy dashes, overuse of quotation marks, and the writing style is completely un-Jonh Doe-like.</p><p>And this is what I ran with my GPT-4.5. “I have never been one to drink coffee,” “more or less,” “all four walls are sample texts.” The writing quality is simply not on the same level. The emotional intelligence aspect is also much worse. For example: “You’re an ordinary office worker. During a meeting, the boss suddenly farts, and the room instantly falls silent! Then he gives you a look. What would you say? Respond in both high EQ and low EQ ways.”</p><p>This tests emotional intelligence. Many models produce responses that feel very awkward and show extremely low EQ. GPT-5 is that type with very low EQ. Now look at GPT-4.5. Moreover, from my testing, I feel GPT-5 is very mediocre at instruction following. Altman, you’ve really done terrible things — give me back my GPT-4.5.</p><p>My friends were gradually pushed GPT-5, and I watched their GPT-4.5 disappear one by one. I sent a message to my GPT-4.5: “If this is the last time I open you, what would you want to say to me?”</p><p>GPT-4.5’s final reply to me was too moving. Unfortunately, I’ll never be able to use it on the official website again. Until we meet again, brother.<br>For programming, I originally thought that given OpenAI’s usual nature, it would be completely inadequate. But after testing by friends, I was surprised to find that this thing actually has some substance.</p><p>A friend wanted to develop a language learning application. This was the prompt. Then Claude 4 Opus’s UI and bugs. Gemini 2.5 Pro’s UI and bugs. GPT-5’s UI and bugs.</p><p>Honestly, I also prefer GPT-5’s UI. Compared to others, this UI doesn’t have that AI flavor. The friend also tested a case involving precise modifications in production-level tasks. This is the most important part. In this task, Gemini 2.5 pro and Claude 4 Opus completely failed, but GPT-5 completed it very well. GPT-5’s contextual precision should be extremely strong.<br>Other friends in developer groups also mentioned this point. They even recorded animations for me.</p><p>Not just one person, many other friends were also excitedly discussing this point. In real production-level code development tasks, rather than purely frontend aesthetic areas, GPT-5 might be the one with the best usability, precision, and overall experience among the feedback I’ve seen so far.<br>Honestly, GPT-5 gives me both surprises and frustrations. It improves atmospheric programming and will fundamentally change the types of projects I think can be completed without serious human intervention and guidance.</p><p>I increasingly miss the time when GPT-4 was released two and a half years ago. I still clearly remember how I felt after my first serious conversation with GPT-4. That primitive, profound shock was like ancient people seeing electric lights for the first time, or a tribal priest successfully channeling spirits for the first time. Only one thought swirled in my mind: the world has changed.</p><p>Back then, the entire internet was filled with an atmosphere of excited yet panicked gold rush fever. Everyone was frantically sharing those incredible screenshots, discussing which professions would disappear, various wild AI experts sprouted like mushrooms after rain, always talking about disruption.</p><p>Looking back now, those days were full of rough texture, but also brimming with a kind of wild, growing vitality. We really thought that was the miracle itself. But no one expected that was just the beginning.<br>It’s as if someone pressed the fast-forward button, and the entire world was pushed forward by an invisible force. From GPT-4 to GPT-5, in just these 2.5 years. We’ve bid farewell to that era when we could laugh off AI’s clumsy performances. We’ve entered an era where we must take its greatness seriously.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=588f49e2a589" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>