Ralph Prompt Audit

Repo-specific prompt drift checker for Ralph workflows.

This skill uses a hybrid approach:

1. Deterministic context collection (aaa completion table --format json + file/example extraction)
2. LLM-as-judge to detect stale, invalid, or misleading command examples

Scope

Audit only these files unless user asks otherwise:

• context/workflows/ralph/**/*.md
• .claude/skills/ralph-*/SKILL.md

Do not normalize slash-command examples (/ralph-plan ..., /ralph-review ...) unless explicitly requested.

Arguments

• check (default): audit and report findings only
• fix: audit and apply high-confidence fixes
• --scope <glob>: optional narrower scope for targeted audits

Workflow

1) Collect deterministic context

Run:

bash .claude/skills/ralph-prompt-audit/scripts/collect-context.sh

This creates tmp/ralph-prompt-audit/ with:

• command-table.json - live CLI matrix from aaa completion table --format json
• ralph-command-matrix.md - condensed command/option view for aaa ralph
• review-files.txt - prompt/skill files to audit
• cli-invocations.txt - extracted aaa ralph examples with file:line
• README.md - generation metadata

2) Run LLM judge pass

Launch a fresh subagent (general) and give it:

• tmp/ralph-prompt-audit/ralph-command-matrix.md
• tmp/ralph-prompt-audit/cli-invocations.txt
• target files from tmp/ralph-prompt-audit/review-files.txt

Ask for strict output in JSON with this schema:

{
  "findings": [
    {
      "file": "path/to/file.md",
      "line": 123,
      "example": "aaa ralph ...",
      "issue": "why this is stale/invalid",
      "suggested": "aaa ralph ...",
      "confidence": "high|medium|low",
      "autoFixSafe": true
    }
  ]
}

Judge criteria:

• command exists in current aaa ralph table
• flags shown are valid for that command
• required flags are present in examples when relevant
• examples do not combine mutually exclusive source flags
• preserve intent and surrounding docs context

3) Apply fixes (only in fix mode)

Apply only autoFixSafe: true findings with high confidence.

If ambiguity remains, keep file unchanged and report it under "needs decision".

4) Verify after changes

Run collector again, then run targeted drift checks (rg) for known stale patterns.

Report:

• files checked
• files changed
• unresolved findings

Repo-specific invariants to enforce

• aaa ralph plan tasks uses --supervised / --headless (not --auto)
• aaa ralph review stories and aaa ralph review gap stories use --milestone <path>
• aaa ralph subtasks complete examples should include --commit <hash> and --session <id>
• aaa ralph plan subtasks source flags are exclusive (--milestone|--story|--task|--file|--text|--review-diary)