Prompt Audit

---
description: Discover and review LLM prompts in this codebase against the best-practices rubric. Reports findings and proposed diffs in the terminal — never edits without approval.
argument-hint: "[path or scope, e.g. ./src or 'changed files']"
allowed-tools: Read, Glob, Grep, Agent
---

You are orchestrating a prompt audit. Findings print to the terminal only — do not
write reports to disk, and do not edit any prompt without explicit user approval.

Scope for this run: **$ARGUMENTS**

Follow these steps:

1. **Resolve scope.** If `$ARGUMENTS` is empty, ask the user what to scan: the whole
   repo, a specific path, or only files changed vs the main branch. Do not proceed
   until scope is clear.

2. **Discover.** Delegate to the `prompt-discovery` subagent over the chosen scope.
   Present the returned inventory as a table and ask the user to confirm or prune it
   (this is the false-positive filter). Drop low-confidence hits the user rejects.

3. **Review in parallel.** For the confirmed inventory, delegate to `prompt-reviewer`
   subagents — batch the prompts and run multiple reviewers concurrently so each works
   in its own context. Pass each the file:line references to review.

4. **Aggregate.** Combine the findings into one terminal report, sorted by severity:
   - A summary line: N prompts reviewed, counts of HIGH/MEDIUM/LOW findings.
   - Per-prompt findings with rubric ids and proposed diffs.
   - A prioritized "fix these first" shortlist.

5. **Offer next step.** Ask whether the user wants to apply any proposed diff (hand off
   to `/prompt-fix`) — do not apply anything automatically.

Always load the `prompt-best-practices` skill so your framing matches the reviewers'.
If `.prompt-review.yaml` exists, mention which overrides are in effect.