[Claude] Improve AI agent setup experience

About a year ago, I made a dumb commitment to myself: build one Claude AI agent or skill per week, every week. Don't blog about it. Don't make YouTube videos about it. Actually build working things and put them on GitHub. I've been doing product management for 30 years — launched over 115 products across my own companies and consulting work. I figured if I'm going to have opinions about AI in product, I should probably understand how it actually works from the inside. Some of what I built: * LegalAnt — a Claude agent for legal teams. Contract review, clause extraction, compliance flagging. Built it because a client was paying a paralegal 3 hours a day to do work that took the agent 4 minutes. It's not perfect. It flags things conservatively and sometimes over-indexes on boilerplate. But it doesn't miss things, which is the actual job. * Market Research agent — structures competitive intelligence work. Maps categories, separates signal from noise, and outputs evidence-graded research briefs. The grading part matters more than people expect. "Here's what I found" is useless. "Here's what I found, and here's how confident you should be in it" is actionable. Most of these were small. Some were bad. A few I deleted and rewrote from scratch. That's the point. Then I built Lumen, which is the big one. Lumen is a Claude Code plugin. 18 agents. 6 end-to-end PM workflows. Runs entirely in your terminal. Before anyone says it — yes, I know. "Another AI PM tool." I was sceptical of my own idea for a while. Here's what made me build it anyway. Every AI PM tool I've tried has the same architecture: you talk to a chatbot, it gives you output, you paste more context, and it gives you more output. You're doing all the coordination in your head. The AI is just an autocomplete with better grammar. What I wanted was something that could actually sequence work. You give it a problem, it figures out which agents need to run in which order, what data each one needs, and what decisions require a human before continuing. More like a junior analyst team than a chatbot. How it actually works: You type something like: /lumen:pmf-discovery Product: \[your product\] Segments: \[your user segments\] Key question: D30 retention dropped from 72% to 61% over 8 weeks. Is this PMF regression, product quality, or both? And it sequences: * EventIQ validates your event schema * SignalMonitor scores PMF by segment from PostHog data * DiscoveryOS builds an opportunity tree from your signals * MarketIQ maps competitive position * DecideWell structures the final decision with evidence weighting Every recommendation gets an evidence quality rating — HIGH / MEDIUM / LOW — based on what data was actually available. If PostHog isn't connected, the PMF scoring step tells you that instead of hallucinating a number. The part I'm most proud of and that sounds the most ridiculous: Each agent is a Markdown file. That's it. YAML frontmatter for config. Markdown sections for behavior. No compiled code. No proprietary framework. If you can write a good product spec, you can write a Lumen agent. Agents talk to each other through named "context slots" — 51 of them defined in a single schema file. An agent either has the slots it needs or it blocks and says what's missing. This made debugging actually possible, which I did not expect. **What's broken / what I'd do differently:** * The setup experience is rough. Getting MCP servers connected requires patience and some comfort with config files. I'm working on this. * 18 agents sounds impressive until you realize some of them are narrow enough that most workflows won't hit them. Enterprise tier agents, especially. * The evidence quality ratings are only as good as the data connected. Without PostHog, W1 is running on vibes with a label on them. * I built this for Claude Code specifically. It won't work in Claude chat. That's a real constraint that I underestimated how much it would limit the audience. Free to start. MIT License. Open on GitHub. I'll keep building one thing a week. Some weeks it's a small skill. Some weeks it's an agent. Occasionally, something bigger. The goal was always to learn in public and share what works. Happy to answer questions about the architecture, what broke, or why I made specific decisions. AMA basically.