Feedback process split agents md vs prompt

When drafting a launch prompt for an agent (Codex or Claude) working in a repo with an AGENTS.md/CLAUDE.md, do not stuff process/workflow instructions (GitHub-boss flow, human merge/deploy gates, dev-slot claiming, webpack-not-turbopack, browser-testing norms, prod-isolation rules) into the prompt itself. Check AGENTS.md first, since most of it is likely already codified there.

Why: Justin caught this directly ('most of this now documented in agents.md or not... it shouldn't need to come from the prompt'). An audit confirmed nearly everything drafted into a 90-line prompt was already in CreatorTrack's AGENTS.md. Only genuinely new structural rules (e.g. a new 'ship multi-phase work as sequential PRs' policy) belong promoted into AGENTS.md as a standing rule.

How to apply: prompt = thin launcher (what to read, what to build, approved decisions, order). AGENTS.md/CLAUDE.md = the how (process, safety, verification). Handoff doc = the what (product spec). When a new structural pattern emerges from a session, promote it into the standing doc rather than repeating it in every future prompt for that repo.

[auto-memory session 577603bc-d3a9-4b19-bb8e-6db80243b054, confidence 0.85, mode direct]