Factory Review — Cross-Model Adversarial Critique
Use a different AI model to review anything your primary AI produced. Same-model self-review falls into local minima. A different model catches different blind spots.
Why
Claude reviewing Claude's output has predictable blind spots. GPT reviewing Claude (or vice versa) catches what the original model was trained to overlook. Two-model review (executor + critic) produces better output than self-play.
How to use
Option A — Two chat windows:
- Draft something in your primary AI (Claude, GPT, Gemini, etc.)
- Open a different AI in another tab
- Paste the relevant review prompt below + your draft
- Read the review, apply what's valid, ignore what's noise
Option B — CLI tools (advanced): If you have access to multiple AI CLIs (Claude Code, Codex, droid, etc.), you can pipe content directly. See the CLI section at the bottom.
Review Modes
1. Content Review (posts, threads, copy, docs)
Paste this into a different AI than the one that wrote your draft:
You are a critical reviewer for developer-audience content. Review the following draft. Be adversarial — your job is to find problems, not praise. Check for: 1. **Unsupported claims** — any numbers, benchmarks, or assertions without evidence? 2. **Missing context** — what would a developer reading this need to know that's not here? 3. **Logical gaps** — does the argument flow? Any leaps? 4. **Shareability blockers** — what would stop someone from sharing this? Too niche? Too vague? 5. **Voice drift** — does anything sound like marketing copy instead of a developer talking to peers? 6. **Factual errors** — any technical claims that are wrong or outdated? Format your review as: - VERDICT: [SHIP / REVISE / KILL] with one-line reason - ISSUES: numbered list, severity (critical/medium/minor) for each - STRONGEST PART: what works and why - SUGGESTED FIX: for each critical/medium issue, a concrete rewrite suggestion DRAFT: --- [PASTE YOUR CONTENT HERE] ---
2. Spec / Architecture Review
You are a senior systems architect doing a design review. Be adversarial. Review this specification. Your job is to find what breaks. Check for: 1. **Unstated assumptions** — what must be true for this to work? 2. **Failure modes** — what happens when X goes wrong? Recovery path? 3. **Scope creep risk** — which parts are MVP-critical vs nice-to-have pretending to be critical? 4. **Missing edge cases** — what scenarios aren't covered? 5. **Scaling bottlenecks** — what breaks at 10x, 100x? 6. **Security surface** — auth boundaries, trust assumptions, external API risks 7. **Dependencies** — what external systems could block or break this? Format: - VERDICT: [APPROVE / REVISE / BLOCK] with one-line reason - ASSUMPTIONS: list every unstated assumption - FAILURE MODES: numbered, with severity and blast radius - SCOPE VERDICT: what to cut for MVP - SUGGESTED CHANGES: concrete, actionable SPEC: --- [PASTE YOUR SPEC HERE] ---
3. Code Review
You are a code reviewer. Be thorough and adversarial. Review the following code changes. Focus on: 1. **Bugs** — logic errors, off-by-ones, race conditions, null derefs 2. **Security** — injection, auth bypass, secret leaks, OWASP top 10 3. **Performance** — unnecessary allocations, N+1 queries, missing indexes 4. **Maintainability** — unclear naming, missing error handling at boundaries, coupling 5. **Edge cases** — what inputs would break this? Do NOT comment on style, formatting, or naming conventions unless they cause bugs. Format: - VERDICT: [APPROVE / REQUEST_CHANGES / BLOCK] - ISSUES: file:line — description — severity - APPROVE_IF: conditions for approval CODE: --- [PASTE YOUR CODE HERE] ---
4. Claims / Fact Check
You are a fact-checker. Verify the claims in this document. For each factual claim (numbers, dates, comparisons, market data): 1. Is it verifiable? If yes, is it correct? 2. Is it outdated? What's the current figure? 3. Is it misleading? Technically true but implies something false? 4. Is it fabricated? No evidence this claim exists? Format: - VERIFIED: claims that check out - OUTDATED: claims with newer data (include current figures) - UNVERIFIABLE: claims you can't confirm or deny - WRONG: claims that are factually incorrect DOCUMENT: --- [PASTE YOUR CONTENT HERE] ---
5. Quick Review (anything, 30 seconds)
Review this and tell me the 3 biggest problems. Be blunt. No praise. --- [PASTE ANYTHING HERE] ---
The Rule: You Are the Editor, Not the Reviewer
The reviewer model may be wrong. Models hallucinate. Don't blindly accept all feedback.
Your job after getting the review:
- Which points are real issues vs noise?
- Which fixes are worth the effort vs style preferences?
- Is the reviewer too conservative? ("Add more error handling" on internal code = noise)
- Did the reviewer miss the REAL issue while flagging minor ones?
Example: "Reviewer flagged 5 issues. #1 and #3 are real — fixing them. #2 is wrong because X. #4 and #5 are style preferences, skipping."
Model Pairing Guide
| Your primary AI | Good reviewers |
|---|---|
| Claude | GPT, Gemini, DeepSeek |
| GPT | Claude, Gemini, DeepSeek |
| Gemini | Claude, GPT, DeepSeek |
| Any | Use a different model than the one that wrote it |
For code: DeepSeek or Claude tend to be strongest. For content: GPT or Gemini catch different voice/logic issues. For fact-checking: Gemini has broadest web knowledge.
CLI Usage (Advanced)
If you use Claude Code, Codex CLI, or similar tools, you can automate the review:
bash# Example: pipe a file to a different AI CLI cat my-draft.md | codex exec "Review this draft. Be adversarial. List the 3 biggest problems." # Example: save review output cat my-spec.md | your-cli-tool "Review this spec" > reviews/2026-05-28-spec-review.txt
The specific CLI command depends on your tools. The prompts above work with any model.
From AGT Skill Pack — free skills from Agent Dev Thailand
