Prompts¶
Every prompt that drives the autonomous work on this project lives here, for transparency and audit. Four categories:
| Category | Source | What it controls |
|---|---|---|
| Critic personas | ~/.claude/skills/prd-critic-loop/prompts/p<N>-*.md |
The 4 PRD critic-loop passes (Staff Engineer, SRE, Security, Codex) — each is a self-contained instruction set the persona AI reads when it runs against a PRD |
| Ralph iteration | scripts/ralph-ci.sh heredoc |
The prompt every ralph iteration's claude --print invocation receives — it tells the iteration AI to read prd.json, find the next story, implement it, run local checks, commit |
| PRD authoring | snarktank/ai-dev-tasks/create-prd.md |
The template AI follows when drafting a fresh PRD from a feature description |
| Session decisions | Curated log of Craig's directional prompts in conversation | Major workflow decisions, scope changes, autonomy authorizations — the "human-in-the-loop" prompts that shaped what the autonomous work has been doing |
Why this matters¶
The autonomous work runs against a few load-bearing prompts. If a PRD is BLESSED, it was BLESSED by these exact persona prompts. If a story landed via ralph, it landed because Claude read this exact ralph-iteration prompt + the AGENTS.md it references. If a major decision was taken (e.g., "use GitHub private not Forgejo"), it was Craig's prompt that drove it.
This page set lets a future reader (or Craig himself, weeks later) reconstruct what the AI was instructed to do, separately from what the AI produced.
How updates work¶
The first three pages are auto-regenerated by update.sh — they read the live source files. If you edit the persona prompts or the ralph script, the next site refresh picks the changes up.
The session-decisions page is curated — Claude updates it after every meaningful directional message from Craig. It is the only Prompts page that is not derived from a deterministic source.