Feedback process belongs in agents md not prompts
name: Standing process lives in AGENTS.md/CLAUDE.md, not in one-off launch prompts description: Justin corrected an over-stuffed 90-line agent prompt: standing 'how to work' process belongs in the repo's AGENTS.md, not repeated in every project prompt or handoff. type: feedback
When drafting a launch prompt for an autonomous coding agent (Codex, Claude worker, etc.), do not re-explain standing process (git workflow, human merge/deploy gates, dev-slot claiming, browser-testing verification norms, safety/isolation rules) inline in the prompt. That process should live once in the target repo's AGENTS.md (or equivalent steering file) and the prompt should be a thin launcher that points at it plus the task-specific handoff/spec.
Why: Justin flagged this directly after I drafted a 90-line prompt duplicating rules already standing in creatortrack/AGENTS.md: "most of this now documented in agents.md or not... a process that doesn't need to come from the prompt." Repeating process in prompts causes drift (prompt and AGENTS.md can silently disagree, e.g. one says phased-PR shipping, the other still says hold-one-preview-then-/ship) and bloats every future prompt.
How to apply: before writing a launch prompt for any repo with an AGENTS.md/CLAUDE.md, audit what's already standing there first. Only add to the prompt what's genuinely new/task-specific (the spec/handoff reference, task-specific phase order, task-specific risk flags). If a process rule is missing and would apply to future projects too (e.g. "large work ships as sequential reviewable PRs, not one mega-branch"), promote it into AGENTS.md itself rather than parking it in the prompt.
[auto-memory session 577603bc-d3a9-4b19-bb8e-6db80243b054, confidence 0.85, mode direct]