An AI agent is a product. The prompt is implementation. Confusing the two is why most agent projects die in production. This repo is the framework for doing it right.
If you read one thing here, read The Agent Product Spec — a seven-section framework for treating an AI agent as a product, not a prompt.
When you're ready to apply it, clone the fillable template and fill it in with your PM, eng lead, and named eval owner in the same room. To run that room well, follow the 90-minute workshop playbook.
claude-code-for-pm — about working with an AI partner as a PM (the operating model). This repo is about building AI agents as products.
Short pieces, each arguing why one section of the spec is the one that decides whether an agent survives production.
- You don't have a prompt problem. You have a Section 1 problem. — on §1, the job to be done.
- The single most overlooked line in an agent spec — on §3, the handoff sentence.
- If engineering owns your agent's eval suite, engineering owns your product — on §5, evaluation criteria.
- Sunset criteria aren't a kill switch. They're a portfolio strategy. — on §7, sunset criteria.
And one worked decision: Agent Platform vs Raw Foundation API — a 5-question lens for the build-surface choice.
And one retrospective: What I Cut Between Spec and Production — five honest cuts, and which ones came back to bite §1.