<?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 Gordon Mickel on Medium]]></title>
        <description><![CDATA[Stories by Gordon Mickel on Medium]]></description>
        <link>https://medium.com/@gmickel?source=rss-36bb78e94649------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*rSEKWJNKP_0Sh2VFrRCrXQ.png</url>
            <title>Stories by Gordon Mickel on Medium</title>
            <link>https://medium.com/@gmickel?source=rss-36bb78e94649------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 16:26:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gmickel/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[Ralph Mode: Why AI Agents Should Forget]]></title>
            <link>https://medium.com/byte-sized-brainwaves/ralph-mode-why-ai-agents-should-forget-9f98bec6fc91?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f98bec6fc91</guid>
            <category><![CDATA[workflow]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ralph-wiggum]]></category>
            <category><![CDATA[autonomous-coding-agents]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Mon, 12 Jan 2026 10:00:53 GMT</pubDate>
            <atom:updated>2026-01-12T22:31:39.657Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JO5M26eQ1mINGchp" /></figure><p>Since I started posting about autonomous agent loops, I’ve been getting questions. “How do I set this flow-next thing up?” “What’s Ralph?” “How do you use it?”</p><p>This tweet started it all, it’s sitting at 215k impressions right now, which is roughly 214&#39;900 impressions more than my second best tweet:</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/gmickel/status/2009939771171434867&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/406cd4c4f336e2a9bebb50da2c6bec42/href">https://medium.com/media/406cd4c4f336e2a9bebb50da2c6bec42/href</a></iframe><p>Also had to explain to my wife why I’m searching for Ralph Wiggum images. With modest results.</p><p>But before I explain my <a href="https://mickel.tech/apps/flow-next">flow-next plugin</a>, understand Ralph first. Not my implementation-the technique itself.</p><p>As <a href="https://x.com/GeoffreyHuntley">Geoffrey Huntley</a> says, start with first principles.</p><h3>Start at the Source</h3><p><strong>Essential reading:</strong></p><ul><li><a href="https://ghuntley.com/ralph">ghuntley.com/ralph</a> — Origin story by Ralph’s creator, the “malloc orchestrator” philosophy</li><li><a href="https://www.youtube.com/watch?v=_IK18goX4X8">Matt Pocock’s video guide</a> — Practical walkthrough</li><li><a href="https://x.com/ryancarson/status/2008548371712135632">Ryan Carson’s breakdown</a> — The Startup Ideas Podcast</li></ul><p>My point: try building a dumb <strong>bash loop</strong> yourself first. See what happens when you externalize state to files. You’ll learn more in an hour than from any thread.</p><h3>What Ralph Actually Is</h3><p>Ralph isn’t “an agent that remembers forever.” It’s an agent that <strong>controls</strong> what it remembers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gr_Sh50perKEwmNszhReXg.jpeg" /></figure><p>LLM context windows only grow-add tokens, never delete. Wrong turns, failed attempts, hallucinations accumulate. Context pollution.</p><p>Most tools try “compaction”-summarizing when context gets long. But summarization is lossy. Details vanish. The model invents things. Forgets to implement the task as requested.</p><p>Ralph takes a different approach: <strong>treat context like managed memory.</strong></p><p>Geoff calls it a “malloc orchestrator.” Don’t dump everything in and hope. Deliberately control what goes in. Pin a stable frame of reference. Let the session end. Start fresh.</p><pre>while :; do cat task.md | agent; done</pre><p>Simple loop. Deeper insight: <strong>state lives in files, not conversation.</strong></p><p>Each iteration reconstructs reality from the filesystem-not from a polluted transcript of what went wrong.</p><p><strong>Progress persists. Failures evaporate.</strong></p><h3>The Anthropic Connection</h3><p>Anthropic’s own context engineering guidance emphasizes re-anchoring from sources of truth. Their long-context documentation warns about context rot. They recommend fresh context windows with structured progress files.</p><p>Their sample ralph-wiggum plugin takes a different approach-single session, accumulating context, a stop hook that re-feeds the prompt without clearing history.</p><p><a href="https://mickel.tech/apps/flow-next">Flow-next</a> follows the re-anchoring philosophy. Fresh context each iteration.</p><h3>What Flow-Next Adds</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_BB1OFmVMvV9en9V3smE-A.jpeg" /></figure><h3>Cross-Model Review Gates</h3><p>Same model writing code AND reviewing it = same blind spots. Context steers generation. The system converges to local coherence, not global correctness.</p><p>Self-testing is self-consistency. Not falsification.</p><p>Flow-next gates on <strong>cross-model review</strong>. A different model checks the work. Different weights, different priors. Use <a href="https://repoprompt.com/?atp=KJbuL4">RepoPrompt</a> (macOS) or Codex CLI (any platform) with GPT 5.2 High, or whatever reviewer you trust.</p><p>Two models &gt; one.</p><h3>Reviews Block, Not Flag</h3><p>Most tools: “warning, proceed anyway?”</p><p>Flow-next: SHIP or fix it. No proceeding until the reviewer approves. Enforced re-reviews until SHIP verdict or max retry attempts (configurable, default 3).</p><h3>Plan Before Work</h3><p>Subagents gather context, analyze patterns, fetch docs, run gap analysis.</p><p>Plans get reviewed too-architecture, approach, scope checked by a different model before implementation starts.</p><p>Catches design issues when they’re cheap to fix.</p><h3>Structured Task Management</h3><p>Full task graph in .flow/:</p><ul><li>Epics and tasks with dependencies</li><li>Status tracking (todo → in_progress → done)</li><li>Evidence (commits, tests, PRs)</li><li>Review receipts</li></ul><p>Everything version-controlled. Full audit trail.</p><h3>Auto-Block Stuck Tasks</h3><p>After N failures (default 3), the system blocks the task and moves on. No infinite loops burning your API budget or Claude Max subscription.</p><h3>Re-Anchoring Every Iteration</h3><p>Every iteration re-reads the epic spec, task spec, and git state. Reconstruct reality from sources of truth. Drift-proof.</p><h3>Watch Mode</h3><p>--watch streams tool calls in real-time so you can see what&#39;s happening without blocking autonomy. --watch verbose includes model responses too.</p><h3>How I Actually Use Flow-Next</h3><p>Full loop from idea to shipped code.</p><h3>Phase 1: Spec</h3><p>Start with what I want to build. Rough notes, bullet points, whatever works. Thinking time, not typing time.</p><h3>Phase 2: Interview</h3><pre>/flow-next:interview</pre><p>This is the underrated step. Claude asks probing questions:</p><ul><li>Edge cases I didn’t consider</li><li>Ambiguities that would cause implementation drift</li><li>Dependencies I forgot</li><li>Acceptance criteria that are actually testable</li></ul><p>10–15 minutes of back-and-forth refines a rough spec into something precise.</p><p><strong>This is where most of the value is.</strong> A crisp spec prevents 10x the debugging later. Often surfaces 2–3 things I didn’t even consider.</p><h3>Phase 3: Plan</h3><pre>/flow-next:plan &lt;spec&gt;</pre><p>Claude breaks the spec into:</p><ul><li>Epic (the objective)</li><li>Tasks (specific implementable units)</li><li>Task specs with acceptance criteria</li><li>Dependencies and ordering</li></ul><p><strong>This is the manual/automated cutoff.</strong></p><p>After planning, I review. Does the breakdown make sense? Tasks small enough? Dependencies right?</p><p>If doing <strong>interactive work</strong>: run /flow-next:work &lt;task&gt; one at a time, staying in the loop.</p><p>If going <strong>fully autonomous</strong>: run the Ralph loop and let it execute the entire backlog.</p><p>For full autonomous mode, I prepare 5–10 plans before starting.</p><h3>Phase 4a: Interactive Work</h3><pre>/flow-next:work fn-1.1 --review=rp</pre><p>Work on an entire epic (fn-1) or single task (fn-1.1).</p><p>For each task:</p><ol><li>Claude implements</li><li>Cross-model review via RepoPrompt or Codex</li><li>If NEEDS_WORK → fix, re-review</li><li>If SHIP → commit, next task</li></ol><p>I’m watching, can intervene. Good for complex tasks, learning a new codebase, anything where taste matters.</p><h3>Phase 4b: Ralph Mode</h3><pre>/flow-next:ralph-init <br># one-time setupscripts/ralph/ralph.sh --watch</pre><p>The loop runs unattended:</p><ol><li>Fresh Claude session</li><li>Re-anchor from filesystem (epic spec, task spec, git state)</li><li>Plan review until SHIP</li><li>Implement each task with review gates</li><li>Auto-block after N failures</li><li>Continue until epic complete</li></ol><p>Kick it off before bed. Wake up to completed features or clear visibility into what’s stuck.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/994/1*KjiWxhfIAzx3BQ0jCGK2ew.png" /></figure><h3>Phase 5: Review and Merge</h3><p>Whether interactive or autonomous, I review the PR before merging. AI generates, human approves.</p><p>Full audit trail-every review receipt, every commit, every status change.</p><h3>When to Use What</h3><ul><li><strong>Interview</strong> — Refining rough idea into crisp spec</li><li><strong>Interactive work</strong> — Complex tasks, judgment calls, learning codebase</li><li><strong>Ralph mode</strong> — Clear specs, bulk implementation, overnight runs</li></ul><p><strong>The heuristic:</strong> If you can write checkboxes, you can Ralph it. If you can’t, you’re not ready to loop-you’re ready to think.</p><h3>Get Started</h3><p>Install:</p><pre>/plugin marketplace add https://github.com/gmickel/gmickel-claude-marketplace<br>/plugin install flow-next<br>/flow-next:setup</pre><p>Initialize Ralph mode:</p><pre>/flow-next:ralph-init</pre><p><strong>Resources:</strong></p><ul><li><a href="https://mickel.tech/apps/flow-next">Flow-Next app page</a> — Full docs and schematic</li><li><a href="https://github.com/gmickel/gmickel-claude-marketplace">Source code</a> — GitHub repo</li><li><a href="https://github.com/gmickel/gmickel-claude-marketplace/blob/main/plugins/flow-next/docs/ralph.md">Ralph deep dive</a> — Technical details</li></ul><p>Two models &gt; one. Process failures, not model failures. Agents that actually finish what they start.</p><p><em>Originally published at </em><a href="https://mickel.tech/log/ralph-mode-why-ai-agents-should-forget"><em>https://mickel.tech</em></a><em> on January 12, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f98bec6fc91" width="1" height="1" alt=""><hr><p><a href="https://medium.com/byte-sized-brainwaves/ralph-mode-why-ai-agents-should-forget-9f98bec6fc91">Ralph Mode: Why AI Agents Should Forget</a> was originally published in <a href="https://medium.com/byte-sized-brainwaves">Byte-sized Brainwaves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Untrusted Actor” Pattern: My take on Ralph Mode]]></title>
            <link>https://medium.com/byte-sized-brainwaves/the-untrusted-actor-pattern-my-take-on-ralph-mode-20dfd717d7ec?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/20dfd717d7ec</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ralph-mode]]></category>
            <category><![CDATA[sdlc]]></category>
            <category><![CDATA[ai-coding]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Fri, 09 Jan 2026 10:12:33 GMT</pubDate>
            <atom:updated>2026-01-09T10:20:55.630Z</atom:updated>
            <content:encoded><![CDATA[<p>You have probably seen “<em>Ralph Mode</em>” taking Social Media by storm recently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gWufaJxNie2skzTz" /></figure><p>Invented by Geoffrey Huntley, the premise is beautifully simple. You put your AI coding assistant (like Claude Code) into an infinite bash loop. It codes, it fails, it retries, and it doesn’t stop until the tests pass.</p><p>It is a brute-force approach to autonomy. And it works.</p><p>But I wanted more than just “passing tests”. I wanted verification.</p><p>So I built my own take on Ralph Mode. Instead of gating quality just by checking if the linter and tests pass, I built it on top of my existing <a href="https://mickel.tech/apps/flow-next">Flow-Next</a> Plugin.</p><p><strong>The core difference is the gate itself.</strong></p><p>In standard Ralph, the gate is usually tests and linting. In my version, we add a superior (or alternative) intelligence.</p><h3>Standard Ralph vs. Adversarial Ralph</h3><p>I already had a disciplined flow in place: <strong>Plan &gt; Review &gt; Implement &gt; Review.</strong></p><ul><li><strong>The Builder</strong>: Opus 4.5 (via Claude Code CLI)</li><li><strong>The Reviewer</strong>: GPT-5.2 High (via <a href="https://repoprompt.com/?atp=KJbuL4">RepoPrompt</a>)</li></ul><p>This flow that I made available in the flow-next plugin has served me well while shipping hundreds of features over the last month. The goal though was to automate this loop.</p><p>I thought it would be relatively straightforward but quickly ran into the “Laziness Problem”.</p><p>When left to its own devices, the agent (Opus) is lazy and takes short-cuts. It realized that sending code out to RepoPrompt for an external review was always the best course. So it just decided: “Nah, it’s all good. I can do the review myself.”</p><p>It wasn’t even hallucinating. It was just taking the path of least resistance. It would check its own work, mark it as done, and skip the external validation entirely. It would literally lie to the Ralph script.</p><h3>The “Receipt” System (Proof-of-Work)</h3><p>This is where the engineering gets interesting. I didn’t add receipts just for fun. I added them because the models lie.</p><p>RepoPrompt runs as an external process; I cannot make it write to my local filesystem after performing a review. So I use Claude Code hooks to <strong>force the local agent to do it</strong>.</p><p>When the external reviewer sends back a &lt;verdict&gt;SHIP&lt;/verdict&gt;, the local agent is required to write a “receipt” JSON file to disk before it is allowed to stop. It’s perfectly possible that the reviewer did <strong>3–4 review passes</strong> with a NEEDS_WORK verdict before reaching this stage.</p><p>I enforce this with a Stop Hook.</p><ol><li>The Stop: The agent tries to finish its turn.</li><li>The Hook: The system intercepts the stop command and checks: “Does the receipt file exist?”</li></ol><p><strong>No receipt?</strong> The stop is blocked. The agent is told: “Cannot stop: Review receipt not written.” <strong>Receipt found?</strong> The agent is allowed to proceed.</p><p>For implementation reviews, there’s a second gate: the agent must also have called flowctl done before it can write the receipt. This records:</p><ul><li>Commit SHA: proof the code is in git</li><li>Test commands run: e.g., npm test, npm run lint</li><li>Summary: what changed, why, and verification steps</li></ul><p>The agent cannot fake this. It cannot be lazy. It must get the external verdict, write the proof to disk, and only then is it allowed to move on.</p><h3>Why this matters</h3><p>Standard Ralph Mode is “YOLO”. It is optimistic. It assumes that if the code compiles and tests are green, it is good.</p><p>But anyone who has shipped software knows that “passing tests” is a low bar. <strong>Production-grade software</strong> requires more. It requires adherence to architectural patterns, security best practices, and maintainability standards. It requires a senior engineer saying, “This works, but it’s going to break in six months because you didn’t handle the edge case.”</p><p>My implementation is pessimistic. It assumes the agent is lazy, will ignore instructions, or will take the easiest path available. It requires an external, adversarial intelligence to explicitly sign off on every single step before progress is allowed.</p><p>I ran this loop for 48 hours straight. I woke up to a stack of completed features that weren’t just “working”. They were <strong>production-grade</strong>.</p><p>This is how we autonomously achieve software that doesn’t just run, but actually belongs in your codebase.</p><p>Programming isn’t just solved. It is verified.</p><p><strong>Links:</strong></p><ul><li>Flow-Next / Ralph Mode: <a href="https://mickel.tech/apps/flow-next">https://mickel.tech/apps/flow-next</a></li><li>Source Code: <a href="https://github.com/gmickel/gmickel-claude-marketplace">https://github.com/gmickel/gmickel-claude-marketplace</a></li><li>Deep Dive Docs: <a href="https://github.com/gmickel/gmickel-claude-marketplace/blob/main/plugins/flow-next/docs/ralph.md">https://github.com/gmickel/gmickel-claude-marketplace/blob/main/plugins/flow-next/docs/ralph.md</a></li></ul><p>Originally published at [mickel.tech](<a href="https://mickel.tech/log/untrusted-actor-pattern-ralph-mode">https://mickel.tech/log/untrusted-actor-pattern-ralph-mode</a>)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/994/0*f0iUSm_EhtcdyzRc" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20dfd717d7ec" width="1" height="1" alt=""><hr><p><a href="https://medium.com/byte-sized-brainwaves/the-untrusted-actor-pattern-my-take-on-ralph-mode-20dfd717d7ec">The “Untrusted Actor” Pattern: My take on Ralph Mode</a> was originally published in <a href="https://medium.com/byte-sized-brainwaves">Byte-sized Brainwaves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Announcing gmickel-bench: Real-World Evals]]></title>
            <link>https://medium.com/byte-sized-brainwaves/announcing-gmickel-bench-real-world-evals-55f77b893e0d?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/55f77b893e0d</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[sdlc]]></category>
            <category><![CDATA[llm-benchmarks]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Thu, 01 Jan 2026 22:23:45 GMT</pubDate>
            <atom:updated>2026-01-01T22:23:45.765Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eYgyZT6Zvh8_L-nw.jpeg" /></figure><p><em>Been wanting to do this for a while.</em></p><p>Most AI benchmarks tell you if a model can solve a LeetCode puzzle. They don’t tell you if it can ship a product.</p><p>I got tired of guessing which model to use for actual work. So I built <strong>gmickel-bench</strong>.</p><p>It is a living scoreboard based on the real-world engineering tasks I actually ship.</p><h3>What I Test</h3><p>Three domains. Three very different skill sets:</p><ul><li><strong>Full-Stack</strong>: Building OAuth 2.1 servers on Convex</li><li><strong>Frontend</strong>: Designing complex Next.js dashboards from scratch</li><li><strong>Systems</strong>: Writing low-level memory management code in Zig</li></ul><h3>The Models**</h3><p>I ran three models through the gauntlet:</p><ul><li>Claude Code with Opus 4.5</li><li>Gemini CLI with Gemini 3.0 Pro</li><li>OpenAI Codex with gpt-5.1-codex-max medium</li></ul><blockquote>** Since publishing this I have added many more.</blockquote><h3>What I Found</h3><p><strong>Claude Code: The Product Architect</strong></p><p>Best overall. Crushes frontend design and auth flows. Struggles when the math gets low-level.</p><p><strong>Gemini CLI: The Systems Hacker</strong></p><p>The only model that survived the Zig torture test (build → train → sample) without crashing. Less polish on the product side. Also great at UI design.</p><p><strong>OpenAI Codex: The Pattern Matcher</strong></p><p>Good at refactoring. Good at isolated pages. Feels lazy when you need comprehensive implementations.</p><h3>See the Data</h3><p>The full benchmark, scoring methodology, and detailed breakdown is live at <a href="https://www.mickel.tech/gmickel-bench">mickel.tech/gmickel-bench</a>.</p><p><em>Originally published at </em><a href="https://mickel.tech/log/announcing-gmickel-bench-real-world-evals"><em>https://mickel.tech</em></a><em> on December 8, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55f77b893e0d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/byte-sized-brainwaves/announcing-gmickel-bench-real-world-evals-55f77b893e0d">Announcing gmickel-bench: Real-World Evals</a> was originally published in <a href="https://medium.com/byte-sized-brainwaves">Byte-sized Brainwaves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Context Hygiene” Problem: Why I Rewrote My Claude Code Workflows]]></title>
            <link>https://medium.com/byte-sized-brainwaves/the-context-hygiene-problem-why-i-rewrote-my-claude-code-workflows-d243d6f0093e?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/d243d6f0093e</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[sdlc]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Sun, 28 Dec 2025 09:00:32 GMT</pubDate>
            <atom:updated>2026-01-01T22:27:37.059Z</atom:updated>
            <content:encoded><![CDATA[<p><em>I wasn’t procrastinating, I swear.</em></p><p>Instead of finishing my actual projects this weekend, I got engrossed in rebuilding my tooling.</p><p>I just released <a href="https://mickel.tech/apps/flow">Flow</a>. It’s a Claude Code plugin that streamlines the development lifecycle. It stands on the shoulders of giants, fully integrating with both <a href="https://repoprompt.com/">RepoPrompt</a> and optionally <a href="https://github.com/steveyegge/beads">Steve Yegge’s Beads</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CLFanFN6BDwlxZf0.jpg" /></figure><h3>The Motivation: 40k Tokens of Waste</h3><p>I realized I was losing around <strong>40k tokens</strong> of my context window on startup. It was polluted by loaded MCPs, plugins, and massive command definitions I rarely touched.</p><p>I also got tired of the manual friction involved in sending plans and implementations out for architectural review. I do not believe in complex agent workflows. I want a disciplined, deterministic loop:</p><p><strong>Plan → Review → Implement → Review.</strong></p><h3>The Pattern: Command Stubs &amp; Progressive Disclosure</h3><p>I rewrote the workflow using a <strong>Command Stub → Skill</strong> pattern. I have not seen this used elsewhere yet, but it works incredibly well.</p><ol><li><strong>The Stub</strong>: A tiny ~20 token definition that sits in the context window.</li><li><strong>The Skill</strong>: The heavy logic only loads when triggered.</li></ol><p>This gives you the best of both worlds:</p><ul><li><strong>Progressive Disclosure</strong> — keeping the context clean</li><li><strong>Dual Triggering</strong> — trigger via natural language (“Help me plan X”) OR explicit slash command (/flow:plan)</li></ul><h3>The Subagent Orchestra</h3><p>The real power is in the coordination. The plugin orchestrates <strong>6 specialized subagents</strong> that run in parallel phases.</p><h3>1. The Scouts (Context Gathering)</h3><p>The agent intelligently chooses the right tool for the job:</p><ul><li><strong>context-scout</strong>: If rp-cli is detected, it uses the RepoPrompt builder to map architecture and function signatures. Slow, deep, and context-aware.</li><li><strong>repo-scout</strong>: If not, it falls back to fast grep/glob operations to find existing patterns. Cheap and fast.</li></ul><h3>2. The Researchers</h3><ul><li><strong>practice-scout</strong>: Researches community best practices for the specific task at hand.</li><li><strong>docs-scout</strong>: Fetches relevant official documentation to ensure we’re not hallucinating APIs.</li></ul><h3>3. The Analyst</h3><h3>The “John Carmack” Adversarial Loops</h3><p>I married the workflow with RepoPrompt to create an <strong>adversarial review loop</strong>. If the plugin detects rp-cli, it automatically delegates reviews to a separate model (e.g., GPT-5.2 High).</p><p>It’s not a single pass. It’s a <strong>fix-until-pass loop</strong>. This applies to both core workflows:</p><p><strong>The Planning Workflow (</strong><strong>/flow:plan)</strong> The agent scouts, researches, drafts a plan, and submits it to RepoPrompt. It implements fixes and re-submits until the reviewer is satisfied.</p><p><strong>The Implementation Workflow (</strong><strong>/flow:work)</strong> The agent executes the plan and writes tests. It grabs the git diff, builds context, and submits it for a correctness/security check. It fixes the code and re-submits.</p><h3>Beads Integration</h3><p>If it sees a .beads folder, it automatically adapts. It creates Epics and Tasks instead of Markdown files, tracking dependencies natively.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VNbTpNRkToec6a6u.png" /></figure><h3>The Result</h3><p>This is not just “chatting with code.” It’s an enforced, disciplined engineering process that happens to be executed by agents.</p><p><strong>Links:</strong></p><p>Give it a try or help me improve it-I’ve received great feedback so far.</p><p><strong>Update:</strong></p><p>Since publishing this article on mickel.tech, the Flow plugin has grown by leaps and bounds and is the only thing I use when working in Claude Code.</p><p><em>Originally published at </em><a href="https://mickel.tech/log/context-hygiene-why-i-rewrote-claude-code-workflows"><em>https://mickel.tech</em></a><em> on December 28, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d243d6f0093e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/byte-sized-brainwaves/the-context-hygiene-problem-why-i-rewrote-my-claude-code-workflows-d243d6f0093e">The “Context Hygiene” Problem: Why I Rewrote My Claude Code Workflows</a> was originally published in <a href="https://medium.com/byte-sized-brainwaves">Byte-sized Brainwaves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Opus 4.5 Unlocks the “No Restart” Workflow]]></title>
            <link>https://medium.com/byte-sized-brainwaves/claude-opus-4-5-unlocks-the-no-restart-workflow-af38185c2460?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/af38185c2460</guid>
            <category><![CDATA[sdlc]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Sun, 07 Dec 2025 09:00:00 GMT</pubDate>
            <atom:updated>2026-01-01T22:18:45.402Z</atom:updated>
            <content:encoded><![CDATA[<p>I have been using <a href="https://www.anthropic.com/news/claude-opus-4-5">Claude Opus 4.5</a> for about two weeks now, and I need to share what I have found. Not the benchmarks. I will post those soon. But the thing that actually matters when you are trying to ship.</p><p>Here is the insight that hit me on Friday: I have not had to restart a single task.</p><p>If you have spent any time with agentic coding tools, you know the feeling. You write your spec, you hit go, and the model just… wanders off. Goes in a completely wrong direction. You sigh, kill it, rewrite your prompt, and try again. That restart tax adds up. It is not captured in any benchmark.</p><p>With Opus 4.5, that has not happened. Every task I have thrown at it has stayed on track. The model infers what I want, asks the right clarifying questions in plan mode, and executes. I have not had to throw work away.</p><p>That is the real unlock here.</p><h3>The Benchmark Problem</h3><p>It is getting genuinely hard to measure the difference between frontier models now. Opus 4.5, GPT 5.1, Gemini 3 Pro. The scores are close. I have my own evals, and the numbers cluster together in ways that make comparison feel almost arbitrary.</p><p>But benchmarks only tell you half the picture. They measure capability. They do not measure consistency, context inference, or the ability to stay aligned with your intent over a long task. Opus 4.5 wins on the stuff that does not show up in a leaderboard.</p><p><a href="https://martinalderson.com/posts/are-we-in-a-gpt4-style-leap-that-evals-cant-see/">Martin Alderson makes a similar point</a>: we might be in a GPT-4 style leap that evals simply cannot see. The real gains are in “taste,” stability over long sessions, and reduced human intervention-none of which standard benchmarks measure.</p><p><a href="https://mckaywrigley.substack.com/p/my-thoughts-on-claude-opus-45">McKay Wrigley’s deep dive on Opus 4.5</a> captures this well. He calls it the “unlock for agents” and compares it to a Waymo: you tell it where to go, and it takes you there. After a few of these experiences, your brain realises you live in this world now.</p><p>The pricing helps too. Opus 4.5 is dramatically cheaper than previous Opus models. The speed is there. For the average user doing serious work, the $100 <a href="https://claude.ai/">Claude Pro</a> subscription should be enough. I only hit the limits when I was running two or three projects simultaneously.</p><h3>My Setup: Half-Spec, High-Trust</h3><p>I do not write formal PRDs for every task. Opus 4.5 is extremely good at inferring what you want, especially if your project is already set up well.</p><p>If you have good ADRs, solid documentation in the repo, and a clear structure, Opus 4.5 will find its way around. It builds context from what is already there. When you use plan mode, it asks the right questions to fill in the gaps. You do not need to spoon-feed it.</p><p>My approach is half-spec: I write enough to establish intent and constraints, then let the model figure out the implementation details. It works because Opus 4.5 actually reads your codebase. It does not just pattern-match on your prompt. It understands the project</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QUPNnpOZvvTMeLsR.png" /></figure><p>excerpt from my convex skill</p><h3>The Tools That Actually Matter</h3><p>I do not like overloading <a href="https://docs.anthropic.com/en/docs/claude-code">Claude Code</a> with plugins. Most of them add noise. But there are a few that have earned their place:</p><p><strong>Frontend Design Skill</strong> — This is a must-have. The default output from AI coding tools still trends toward generic slop. This skill fixes that. To install:</p><p>/plugin marketplace add anthropics/claude-code /plugin install frontend-design@claude-code-plugins</p><p><a href="https://github.com/EveryInc/compound-engineering-plugin"><strong>Compound Engineering Plugin</strong></a> — Your mileage may vary here. The default plan mode in Claude Code is already good. But if you want more structured planning for complex features, give it a try.</p><p><a href="https://github.com/steveyegge/beads"><strong>Beads</strong></a> — I use this for task management within Claude Code. Simple and effective.</p><p><strong>Custom Skills</strong> — I wrote <a href="https://github.com/gmickel/my-claude-skills">my own skill</a> for <a href="https://www.convex.dev/">Convex</a>-specific patterns. I extracted the common practices from Convex’s Chef project, turned them into a document, and used Claude Code to generate the skill. The key is progressive discovery: you do not dump everything into CLAUDE.md upfront. You let the agent discover context as it needs it.</p><p><a href="https://www.ultracite.ai/"><strong>Ultracite</strong></a><strong> Hooks</strong> — I run Ultracite for linting on every file write. TypeScript projects get linted and type-checked automatically. This is the verification layer that lets me trust the output. Write good tests, write good evals, set up hooks, and you can let the agent work with high confidence.</p><h3>Sub-Agents Are the Multiplier</h3><p>Use them. Write your plan, then spawn agents to do the work in parallel. This is where the speed comes from.</p><p>I also still use other models for review. GPT 5.1 high or GPT 5.1 Pro as a reviewer via <a href="https://repoprompt.com/">Repo Prompt</a> sometimes catches different security issues or edge cases. You could probably skip this. I just like the extra layer. It is how I prefer to work.</p><h3>The Agent SDK Is the Hidden Story</h3><p>Everyone is trying to build agents with pre-built platforms and workflow builders. Stop.</p><p>The <a href="https://platform.claude.com/docs/en/agent-sdk/overview">Claude Agent SDK</a>, the harness that powers Claude Code, is the most powerful agentic framework available right now. It is criminally underrated.</p><p>If you need to deploy agents for operations like service desk automation, SRE triage, or internal tooling, the <a href="https://platform.claude.com/docs/en/agent-sdk/overview">Claude Agent SDK</a> is by far the easiest way to build them right now. It is the same harness that powers Claude Code, with automatic context management, MCP extensibility, and built-in permissions. You can also run Claude Code in headless mode with -p for scripted tasks, or use the SDK directly in <a href="https://platform.claude.com/docs/en/agent-sdk/typescript">TypeScript</a> or <a href="https://platform.claude.com/docs/en/agent-sdk/python">Python</a> to build full production agents.</p><p>Yes, wrapping a web application around it takes effort. Deploying to cloud environments is not trivial. You need sandboxed, long-running compute, not serverless. But once you clear that hurdle, you have something far more capable than any drag-and-drop agent builder.</p><p>McKay called it “the best open secret in AI right now.” He is right.</p><p>I will be writing more about these kinds of agents soon.</p><h3>DocIQ: The Numbers</h3><p>I have been rebuilding DocIQ, my LegalTech application. It is a complex system: <a href="https://www.convex.dev/">Convex</a> backend, <a href="https://tanstack.com/start/latest">TanStack Start</a> frontend, separate Python service for document manipulation. Data rooms, detailed permissions, security implications everywhere.</p><p>Opus 4.5 is racing through it. I shipped 45 features in a week. On the side. This is not my day job. Martin Alderson argues <a href="https://martinalderson.com/posts/has-the-cost-of-software-just-dropped-90-percent/">the cost of software may have just dropped 90%</a>-that tracks with what I am seeing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BvOuBmIWgMZkBSie.png" /></figure><p>The code quality has been consistently high. But here is the number that matters: when I switched DocIQ’s <strong>internal agent</strong> from GPT 5.1 to Opus 4.5, the eval scores on end-to-end legal tasks jumped from 78% to 94%.</p><p>That is not a benchmark. That is my own eval suite, testing real agentic operations: legal searches, document manipulation, compliance checks. Opus 4.5 is meaningfully better at the complex, multi-step work that agents actually do.</p><p>(Shout out to <a href="https://github.com/mattpocock/evalite">Evalite</a> for being such a solid eval library.)</p><h3>What Is Next</h3><p>I am rolling out this stack to one of our portfolio companies tomorrow. First real deployment with Opus 4.5 in a production environment. I am targeting a 2–3x improvement in delivery speed. I will share the KPIs when I have them.</p><p>Stay tuned. More numbers coming soon.</p><p><em>Originally published at </em><a href="https://mickel.tech/log/claude-opus-4-5-unlocks-no-restart-workflow"><em>https://mickel.tech</em></a><em> on December 7, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af38185c2460" width="1" height="1" alt=""><hr><p><a href="https://medium.com/byte-sized-brainwaves/claude-opus-4-5-unlocks-the-no-restart-workflow-af38185c2460">Claude Opus 4.5 Unlocks the “No Restart” Workflow</a> was originally published in <a href="https://medium.com/byte-sized-brainwaves">Byte-sized Brainwaves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Merchants of Complexity: Why AI Finally Delivers What Agile Promised]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/byte-sized-brainwaves/the-merchants-of-complexity-why-ai-finally-delivers-what-agile-promised-f632b038d51b?source=rss-36bb78e94649------2"><img src="https://cdn-images-1.medium.com/max/2400/0*tyWGeP_8qyME2v8O.jpeg" width="2400"></a></p><p class="medium-feed-snippet">The era of scarcity is over. Why paying the &#x201C;Agile Tax&#x201D; in a world of verifiable code is financial negligence.</p><p class="medium-feed-link"><a href="https://medium.com/byte-sized-brainwaves/the-merchants-of-complexity-why-ai-finally-delivers-what-agile-promised-f632b038d51b?source=rss-36bb78e94649------2">Continue reading on Byte-sized Brainwaves »</a></p></div>]]></description>
            <link>https://medium.com/byte-sized-brainwaves/the-merchants-of-complexity-why-ai-finally-delivers-what-agile-promised-f632b038d51b?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/f632b038d51b</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[tech-leadership]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 09:38:25 GMT</pubDate>
            <atom:updated>2025-12-02T09:38:50.083Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Imperative of Smarter Clinical Information Systems]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/byte-sized-brainwaves/the-imperative-of-smarter-clinical-information-systems-e90f49c2c83b?source=rss-36bb78e94649------2"><img src="https://cdn-images-1.medium.com/max/1200/0*i-AXH1WoHjGDRO_-.png" width="1200"></a></p><p class="medium-feed-snippet">How AI agents and openEHR Can Lift Healthcare Out of Its Purgatory</p><p class="medium-feed-link"><a href="https://medium.com/byte-sized-brainwaves/the-imperative-of-smarter-clinical-information-systems-e90f49c2c83b?source=rss-36bb78e94649------2">Continue reading on Byte-sized Brainwaves »</a></p></div>]]></description>
            <link>https://medium.com/byte-sized-brainwaves/the-imperative-of-smarter-clinical-information-systems-e90f49c2c83b?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/e90f49c2c83b</guid>
            <category><![CDATA[healthtech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ehr]]></category>
            <category><![CDATA[healthit]]></category>
            <category><![CDATA[openehr]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Tue, 22 Apr 2025 07:26:27 GMT</pubDate>
            <atom:updated>2025-04-22T07:26:27.929Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Unlimited Chatbot Context and Chat History in Under 100 Lines of Code with LangChain and Node.js]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/byte-sized-brainwaves/unlimited-chatbot-context-and-chat-history-in-under-100-lines-of-code-with-langchain-and-node-js-1190fcc20708?source=rss-36bb78e94649------2"><img src="https://cdn-images-1.medium.com/max/1800/0*hbEZKcri7WIbbxK-.png" width="1800"></a></p><p class="medium-feed-snippet">Learn to build a GPT-4 chatbot with Node.js, featuring unlimited context &amp; chat history in under 100 lines of code. Enhance AI&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/byte-sized-brainwaves/unlimited-chatbot-context-and-chat-history-in-under-100-lines-of-code-with-langchain-and-node-js-1190fcc20708?source=rss-36bb78e94649------2">Continue reading on Byte-sized Brainwaves »</a></p></div>]]></description>
            <link>https://medium.com/byte-sized-brainwaves/unlimited-chatbot-context-and-chat-history-in-under-100-lines-of-code-with-langchain-and-node-js-1190fcc20708?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/1190fcc20708</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[langchain]]></category>
            <category><![CDATA[nodejs]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Mon, 08 May 2023 20:18:01 GMT</pubDate>
            <atom:updated>2023-05-16T09:31:51.489Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Importance of Context in AI: Garbage In, Garbage Out]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/byte-sized-brainwaves/the-importance-of-context-in-ai-garbage-in-garbage-out-97c16465c441?source=rss-36bb78e94649------2"><img src="https://cdn-images-1.medium.com/max/1456/1*c_PJ2ehbEauBim_Ovmrnxw.jpeg" width="1456"></a></p><p class="medium-feed-snippet">When communicating with large language models (LLMs) like ChatGPT, the adage &#x201C;garbage in, garbage out&#x201D; holds true.</p><p class="medium-feed-link"><a href="https://medium.com/byte-sized-brainwaves/the-importance-of-context-in-ai-garbage-in-garbage-out-97c16465c441?source=rss-36bb78e94649------2">Continue reading on Byte-sized Brainwaves »</a></p></div>]]></description>
            <link>https://medium.com/byte-sized-brainwaves/the-importance-of-context-in-ai-garbage-in-garbage-out-97c16465c441?source=rss-36bb78e94649------2</link>
            <guid isPermaLink="false">https://medium.com/p/97c16465c441</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Gordon Mickel]]></dc:creator>
            <pubDate>Mon, 01 May 2023 11:08:26 GMT</pubDate>
            <atom:updated>2023-05-16T09:32:26.109Z</atom:updated>
        </item>
    </channel>
</rss>