Have agents co-build a reference system with role splits

A solo founder divided an AI-assisted build into roles and shipped a 120-model mental-models framework with an MCP server.

Done with ClaudeChatGPTCursor· Advanced· Nov 2025

What they did
They set up Claude, ChatGPT, Windsurf, and Cursor as separate collaborators, each with a defined job, and kept them moving in parallel with a SITREP check-in routine. The agents worked from the framework’s own mental models to shape the product, check assumptions, and carry it through implementation. A five-question wickedness rubric was used to sort issues by difficulty. The project used React, Cloudflare Workers, D1, and TypeScript.
How it went
The finished system included 120 models and earned a 9.2/10 quality score. It also passed 140 chaos tests and produced an MCP server for Claude Desktop. The work took 22 months end to end.
Tips
Break the work into clearly named roles so the agents do not overlap too much. Keep a pause point before any public release or other irreversible action.

Try it yourself with Cursor

Act like a small product-and-engineering crew for me. I’m building a [number]-item [topic/framework] as a set of entries, plus an MCP server so my AI client can retrieve one by title, keyword, or scenario. First, define the storage format and project layout; next, write [number] starter entries; then implement the lookup server and tests. Give each phase a different role, share short progress updates as you go, and stop to ask for my approval before any deployment, publication, payment, deletion, or submission. When you think it is ready, confirm that a query from [AI client] returns the expected entry.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on news.ycombinator.com ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.