ai Review Modes

---
id: PROMPT-ai-review-modes
layer: prompt
purpose: The semantic (LLM-side) checklists for each AI-review mode - the judgment half of the contract whose deterministic half is scripts/run_ai_review.py; findings return as JSON validated against schemas/recommendation.schema.yaml via --ingest.
read_when: An LLM session (the /ai-review slash command, the adaptive-harness skill, or any reviewer) is executing a review mode and needs the mode's judgment checklist and output contract.
depends_on:
  - ../scripts/run_ai_review.py
  - ../schemas/recommendation.schema.yaml
  - ../schemas/review_report.schema.yaml
  - ../docs/codex-delegation-policy.md
used_by:
  - SCRIPT-run-ai-review
  - global /ai-review slash command (~/.claude/commands/ai-review.md)
tags: [prompt, ai-review, modes, checklist, semantic-judgment, ingest]
retrieval_keywords: [ai review modes, harness cleanup checklist, invocation efficiency questions, codex delegation review, diff review, experiment review, scheduled review, ingest findings json, keep simplify remove replace experiment]
---

# AI-review modes — semantic checklists

**Contract.** `scripts/run_ai_review.py` does the deterministic work (inventory,
integrity scans, telemetry harvest, report/history writing). This file holds the
judgment questions per mode. An LLM session answers them **with cited evidence**
(file:line, telemetry numbers, incident ids), writes findings as JSON records per
`schemas/recommendation.schema.yaml` (the runner enforces required fields,
enums, the REC id pattern, and the human-approval flag on high-risk
Remove/Replace records — not full JSON-Schema validation), and hands them
back via: