<?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 Ricardo Ruiz on Medium]]></title>
        <description><![CDATA[Stories by Ricardo Ruiz on Medium]]></description>
        <link>https://medium.com/@ruizrica2?source=rss-b25447eddbd2------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*R8WbFdZW7vhrCrDziQ4ZhQ.jpeg</url>
            <title>Stories by Ricardo Ruiz on Medium</title>
            <link>https://medium.com/@ruizrica2?source=rss-b25447eddbd2------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:17:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ruizrica2/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[Updates to toolkit, /agent-plan and /agent-spec
https://github.com/ruizrica/toolkit]]></title>
            <link>https://medium.com/@ruizrica2/updates-to-toolkit-agent-plan-and-agent-spec-https-github-com-ruizrica-toolkit-8cf7f88f2303?source=rss-b25447eddbd2------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cf7f88f2303</guid>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[claude-code-skills]]></category>
            <dc:creator><![CDATA[Ricardo Ruiz]]></dc:creator>
            <pubDate>Sat, 18 Apr 2026 21:56:25 GMT</pubDate>
            <atom:updated>2026-04-18T21:56:25.414Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/784/0*QbmQ_VBdelXZOeE6.png" /></figure><p>Updates to toolkit, /agent-plan and /agent-spec<br><a href="https://github.com/ruizrica/toolkit">https://github.com/ruizrica/toolkit</a></p><p>/agent-plan is for the single-document case. I drop in a rough topic, it runs a quick scan of my repo for context, then asks me structured questions in-line: what’s the goal, what’s in scope, what’s out, what could go wrong, how do we verify it’s done.</p><p>When it has enough, it writes a full plan to .context/plans/, opens it in a browser where I can edit or annotate, and waits for my explicit approval. The moment I approve, execution starts.</p><p>/agent-spec is the same shape, but for larger features that need a requirements/design/tasks split. It follows the Kiro pattern EARS-format requirements, a design doc with Mermaid diagrams, and a task breakdown with test-first hints and gates on a spec review before it starts building.</p><p>Two details matter more than they sound.</p><p>The review is round-trip, not cosmetic. If I edit the plan in the browser, those edits get written back to disk, and execution uses the edited version. I’m not approving a screenshot. I’m approving the artifact.</p><p>The approval is binding. Declining is a real path. If I look at the plan and don’t like it, I click decline, leave a comment, and the command loops back to the assembly step instead of ploughing ahead. No wasted cycles, no half-committed work to unwind.</p><p>The part I didn’t expect to need: the completion viewer</p><p>The piece that genuinely surprised me is what happens after execution.</p><p>When the work finishes, a completion report auto opens in the same browser viewer summary of what shipped, a Mermaid diagram of the change flow, per-file unified diffs, and the task checklist ticked off. All pulled straight from the session and the git diff, no manual write-up, no separate “show me what happened” prompt.</p><p>If this pattern feels familiar, it’s because I’ve been running it for a while in my other project agent-pi, where this round-trip review layer is first-class. I lean on Claude Code hard for day to day work, and Anthropic’s recent restrictions on using Max-plan OAuth against third-party services closed off the path of just plugging agent-pi’s infrastructure into Claude Code directly. So I ported the piece I cared about most the viewer into the toolkit plugin, wired through agent-viewer. Same experience, same round-trip contract, just running natively inside Claude Code instead of across the two.</p><p>Claude Code runs as a CLI you live in the shell. The completion viewer breaks that constraint without breaking the flow: the agent is still the thing driving execution, I’m still in my terminal, but for the few moments when I want to actually look at what a long task produced, a real review surface appears editable, annotatable, and written back to disk if I touch it.</p><p>That closes the full loop: plan in the browser, approve, execute in the terminal, review in the browser again without ever leaving Claude Code. The plan, the code, and the post-mortem are all the same round-trip artifact. Three things I used to keep in three separate places, collapsed into one.</p><p>It’s a small trick architecturally the CLI spins up a local web view and blocks on a result but it changes how much rigor I’m willing to apply on a Tuesday afternoon. When the ceremony costs you nothing, you actually do it.</p><p>Repo: <a href="https://github.com/ruizrica/toolkit">https://github.com/ruizrica/toolkit</a> the commands are /agent-plan and /agent-spec. The original full-stack version lives at <a href="https://github.com/ruizrica/agent-pi">https://github.com/ruizrica/agent-pi</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cf7f88f2303" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Built a Text Cleaning App With Zero Dependencies Because I Was Tired of Pasting Into Random…]]></title>
            <link>https://medium.com/@ruizrica2/i-built-a-text-cleaning-app-with-zero-dependencies-because-i-was-tired-of-pasting-into-random-280b280f020e?source=rss-b25447eddbd2------2</link>
            <guid isPermaLink="false">https://medium.com/p/280b280f020e</guid>
            <dc:creator><![CDATA[Ricardo Ruiz]]></dc:creator>
            <pubDate>Sat, 21 Mar 2026 04:21:00 GMT</pubDate>
            <atom:updated>2026-03-21T04:21:00.452Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*jikF-IQ-NTllmDKdgLo7wg.png" /></figure><h3>I Built a Text Cleaning App With Zero Dependencies Because I Was Tired of Pasting Into Random Websites</h3><p>I clean up text constantly. Removing extra whitespace, sorting lines, fixing casing, stripping duplicates. For years I’d paste stuff into whatever random website showed up first on Google, get blasted with ads and cookie banners, and wonder if my clipboard contents were being shipped off to some analytics pipeline.</p><p>So I built my own. It’s called tex.</p><h3>What It Is</h3><p>Text Tools is a single page web app for transforming text. Paste messy text on the left, click operations from the toolbar, see clean results on the right. No backend. No build step. No npm packages. You clone the repo, open index.html, and you&#39;re done.</p><p>I wanted something I could keep open in a tab all day and reach for without thinking about it. That’s exactly what this turned into.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/503/1*8-NsXyzRI5LrF-qpfVTrvw.png" /></figure><h3>Why I Made It</h3><p>I kept running into the same friction. I’d be cleaning up a markdown file or formatting some standup notes and I’d need to do three or four things in a row: trim whitespace, remove duplicate lines, sort alphabetically, convert to lowercase. Every online tool I found would do one thing at a time, so I’d bounce between tabs copying and pasting over and over.</p><p>I wanted a pipeline. Click a few operations and have them stack on top of each other. Toggle any of them on or off without losing the chain. That’s the core idea.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7gFz7B2F6Ol4DjweWSw8kA.png" /></figure><h3>The Operations</h3><p>There are 15 operations organized into groups:</p><p><strong>Whitespace</strong> — Remove extra spaces, remove extra line breaks, trim whitespace, flatten to a single line.</p><p><strong>Case</strong> — UPPERCASE, lowercase, Title Case, camelCase.</p><p><strong>Lines</strong> — Sort A to Z, sort Z to A, natural sort, remove duplicates.</p><p><strong>Encode</strong> — URL encode and URL decode.</p><p><strong>Search</strong> — Find and replace with support for plain text and regex.</p><p>Nothing exotic. Just the stuff I actually use every day.</p><h3>The Stackable Pipeline</h3><p>This is probably the feature I care about most. When you click an operation it gets added to an effects bar at the bottom. Each one shows up as a toggleable pill. You can click the pill to disable it temporarily or click the X to remove it entirely. Case and sort operations are mutually exclusive within their group so things don’t conflict.</p><p>The result pane updates live as you add or remove effects. It feels like building a little text processing pipeline right in the browser.</p><h3>The Diff View</h3><p>I added an inline diff view because I kept second guessing what had actually changed after running a bunch of transforms. Toggle it on and the result pane highlights removed text in red with strikethrough and added text in green. It’s word level, not line level, so you can see exactly what moved.</p><h3>Design Choices</h3><p>The whole thing runs on vanilla JavaScript, plain CSS, and semantic HTML. No framework. No bundler. No transpiler. Every file stays under 500 lines and each one has a single job:</p><ul><li>index.html — markup and toolbar</li><li>styles.css — dark theme and layouts</li><li>app.js — UI orchestration and state</li><li>operations.js — pure text transform functions</li><li>diff.js — word level diff algorithm</li><li>history.js — in memory undo and restore</li></ul><p>I enforced separation of concerns because I wanted to be able to open any file and immediately know what it does. No hunting through a component tree or tracing imports across six directories.</p><p>The visual style is a dark theme inspired by GitHub’s dark mode. Blue accent shades. Monochrome SVG icons. No emojis. No gradients. No distractions. I spend all day looking at text and I wanted something that got out of the way.</p><h3>No Build Step Is the Feature</h3><p>I’m serious about this. You clone the repo, open the HTML file, and it works. There’s no npm install. There&#39;s no webpack config. There&#39;s no dev server. The only external dependency is Lucide for the toolbar icons, loaded from a CDN.</p><p>I’ve watched too many simple tools balloon into complex projects because someone added a build pipeline early on and then the tooling became its own maintenance burden. I didn’t want that. This is a text cleaning utility. It should be simple to run and simple to modify.</p><h3>If You Want to Try It</h3><pre>git clone https://github.com/ruizrica/text-tools.git<br>cd text-tools<br>open index.html</pre><p>That’s it. MIT licensed. If you work with text regularly and you’re tired of pasting into sketchy websites, give it a shot. It’s small, it’s fast, and it does exactly what it says.</p><p>GitHub: <a href="https://github.com/ruizrica/text-tools">https://github.com/ruizrica/text-tools</a></p><p>Cheers 🍻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=280b280f020e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Turned Pi Into a Multi-Agent Coding Assistant]]></title>
            <link>https://medium.com/@ruizrica2/how-i-turned-pi-into-a-multi-agent-coding-assistant-cb1c2f1c8145?source=rss-b25447eddbd2------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb1c2f1c8145</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-engineering]]></category>
            <dc:creator><![CDATA[Ricardo Ruiz]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 00:29:06 GMT</pubDate>
            <atom:updated>2026-03-25T11:36:22.323Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*qbtdhYaK5jn7Esdu--7DSw.png" /></figure><p>How I Turned Pi Into a Multi-Agent Coding Assistant</p><p>I’ve been using Pi the terminal-based AI coding agent by <a href="http://twitter.com/badlogic">@badlogic</a> as my daily driver for a while now. Out of the box, it’s already a solid single agent assistant with tool use, conversation memory, and a nice TUI. But over time I kept wanting more from it, so I started writing extensions. One turned into five, five turned into 28, and at some point the whole thing became its own project: agent.</p><p>Today I’m open sourcing it. It’s been my private toolkit for a while, but it’s reached a point where I use it every day and I think other people might get something out of it too. So here it is.</p><p><strong>What I Built</strong></p><p>Agent is really just a collection of extensions and config files that sit on top of Pi. No forks, no patches everything loads through Pi’s extension API and a `settings.json` file. The idea was to keep it portable and easy to update when Pi itself gets new releases.</p><p>The biggest additions are six operational modes I can cycle through with Shift+Tab. Most of the time I’m in PLAN mode, which forces me to think through what I’m doing before the agent starts touching files. There’s also SPEC mode for when I need to shape requirements first, and TEAM/CHAIN/PIPELINE modes for when a task is complex enough that it helps to break it across multiple agents with different roles.</p><p><strong>The Multi-Agent Stuff</strong></p><p>This is probably the part I’m most stoked about, though it’s still evolving. Teams, chains, and pipelines are all defined in YAML files. A team is just a list of agent names — a planner, a builder, a reviewer — and the primary agent dispatches work to them. Chains are sequential: one agent’s output feeds into the next. Pipelines combine phases with parallel dispatch.</p><p>Agent definitions are Markdown files with YAML frontmatter. Adding a new specialist is as simple as writing a `.md` file and referencing it in the config. It’s not fancy, but it works well for my workflow.</p><p><strong>Security is Important to Me</strong></p><p>I added a security layer early on because I was nervous about letting agents run commands unsupervised. There’s a pre-tool-hook that blocks dangerous commands before they execute, a content scanner that checks for prompt injection in tool results, and system prompt hardening at session start. There’s also a `/secure` command that can audit a project and install basic protections. It’s not bulletproof, but it gives me peace of mind.</p><p><strong>The Browser Viewers</strong></p><p>This part started as a quick hack and turned into one of my favorite features. There are browser-based GUIs that open right from the terminal for reviewing plans, reading completion reports with diffs and rollback controls, approving specs with inline comments, and browsing past reports. They make the terminal workflow feel a lot more interactive without leaving the environment I’m comfortable in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PiSEBBEZBNHG99H8" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*M1b0YCbw7YO7CAZ2.png" /></figure><p><strong>Models and Themes</strong></p><p>I’ve configured over 30 models across providers like Anthropic, OpenAI, Google, Mercury, Synthetic, and OpenRouter. Anything with an OpenAI-compatible API plugs right in. There is 11 terminal themes because I spend all day looking at this thing and wanted it to look nice. Dark is my current favorite, but I rotate.</p><p><strong>Commander</strong></p><p>There’s a companion project called Commander that acts as a dashboard Kanban board, spec wizard, Jira integration. Agent connects to it over MCP and tasks sync both ways. I use them together, but they’re fully independent. You don’t need Commander to use agent, and vice versa.<br>- Coming Soon to open source.</p><p><strong>If You Want to Try It</strong></p><p>It’s MIT licensed and the setup is pretty simple: Bun, Pi installed globally, clone the repo, `npm install`, and run `pi` from the agent directory. Everything loads automatically from there.</p><p>I’m still actively working on this, so things will change and there’s plenty to improve. But if you’re a Pi user looking to push it further, or just curious about multi-agent orchestration in a terminal-first workflow, feel free to poke around. Contributions are welcome, and I’m happy to answer questions.</p><p>GitHub: <a href="https://github.com/ruizrica/agent-pi">https://github.com/ruizrica/agent-pi</a></p><p>Cheers 🍻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb1c2f1c8145" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[my daily toolkit]]></title>
            <link>https://medium.com/@ruizrica2/my-daily-toolkit-e6bf0e02670f?source=rss-b25447eddbd2------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6bf0e02670f</guid>
            <category><![CDATA[claude-code-tips]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Ricardo Ruiz]]></dc:creator>
            <pubDate>Mon, 23 Feb 2026 06:58:46 GMT</pubDate>
            <atom:updated>2026-02-23T07:02:18.342Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the past year, I’ve been quietly assembling a small set of tools that have become essential to my daily workflow as a software developer.</p><p>These aren’t grand frameworks or polished products they’re the kind of pragmatic, purpose built tools you reach for every single day without thinking about it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/784/0*zYlIs4Y7dC0f-LKO.png" /></figure><h3>toolkit</h3><p>The centerpiece is <a href="https://github.com/ruizrica/toolkit">toolkit</a>, a Claude Code plugin I’ve built up over time. It includes multi-agent orchestration patterns, TDD workflows, and a growing set of productivity commands I’ve refined through actual use not hypotheticals.</p><p>If you use Claude Code regularly, you know how much of your time is spent on repetitive scaffolding. Toolkit was my answer to that. It handles the patterns so I can focus on the problem.</p><p><a href="https://github.com/ruizrica/toolkit">https://github.com/ruizrica/toolkit</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/432/0*IYecmzcoTwWDZG7I.png" /></figure><h3>rust-debug-mcp</h3><p><a href="https://github.com/ruizrica/rust-debug-mcp">rust-debug-mcp</a> is a Model Context Protocol server for debugging and operating Rust Tauri applications over WebSocket. If you’ve ever tried to get real visibility into a running Tauri app, you know the pain. This solves it.<br><a href="https://github.com/ruizrica/rust-debug-mcp">https://github.com/ruizrica/rust-debug-mcp</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/0*5iB_xeu_Yp4PD1jV.png" /></figure><h3>electron-debug-mcp</h3><p><a href="https://github.com/ruizrica/electron-debug-mcp">electron-debug-mcp</a> takes a similar approach for Electron deep Chrome DevTools Protocol integration through an MCP server. It gives Claude real, live access to your running app so debugging becomes a conversation rather than a guessing game.<br><a href="https://github.com/ruizrica/electron-debug-mcp">https://github.com/ruizrica/electron-debug-mcp</a></p><p>I hope these save you some of the time they’ve saved me. Cheers.</p><p>Ricardo<br>ruizrica.io</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e6bf0e02670f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>