blind-review

name	blind-review
description	Single-blind document review and devil's advocate stress testing. Use when reviewing any proposal, PRD, spec, or plan to catch expression gaps and false consensus. Triggers on "blind review", "review this proposal", "stress test", "find holes", "sanity check", "devil's advocate".
when_to_use	When the user has a document or proposal they want critically evaluated. Especially valuable after collaborative AI sessions where shared context may have created blind spots. Also use when multiple reviewers agree too easily.
disable-model-invocation	true
allowed-tools	Read Write Edit Bash

Blind Review

Catch blind spots that context-aware reviewers miss. Uses single-blind evaluation (reviewer doesn't see original intent) and devil's advocate stress testing.

Model compatibility: Gemini is recommended for blind review, but any LLM works — GPT, Qwen, Doubao, DeepSeek, etc. Use a different provider from Claude so biases don't overlap. "Gemini" below is shorthand for your chosen model.

Quick start

/blind-review path/to/proposal.md

Core insight

When a reviewer knows your original idea, they unconsciously fill in gaps you left in the document. A reviewer who only sees the document — with no background — evaluates what you actually wrote, not what you meant to write. The gap between these two is where miscommunication lives.

Step 1: Single-blind review prompt

Read the document the user wants reviewed. Generate a prompt for a brand new Gemini session (no prior context).

The prompt must contain:

• ONLY the document content
• NO original idea, NO background, NO motivation, NO "what I was trying to do"
• Clear evaluation questions

Template:

Here is a document. I have no additional background information for you.
Please evaluate it purely on its own merits.

1. In one sentence, what is this document proposing?
2. Is the logic internally consistent? Where does it "not read right"?
3. As a decision-maker seeing this for the first time, would you approve it? Why?
4. Which parts are ambiguous, vague, or require unstated context to understand?
5. What is the single biggest risk this document ignores?
6. Your improvement suggestions (prioritized).

Document:
---
[paste document content here]
---

After generating, auto-copy to clipboard:

cat prompts/gemini-blind-review.md | pbcopy 2>/dev/null || cat prompts/gemini-blind-review.md | xclip -selection clipboard 2>/dev/null || echo "Prompt saved to prompts/gemini-blind-review.md — please copy manually"

Session management tip: Tell the user to name their browser tab "Blind Review" to avoid mixing it up with other Gemini sessions.

Step 2: Analyze divergence

After the user provides Gemini's response:

1. Compare Gemini's one-sentence summary (question 1) against the document's stated intent

2. Classify the divergence:

   • No divergence → document expression is clear, proceed
   • Minor divergence → phrasing issues, fix specific sections
   • Major divergence → structural problem, may need to revisit the concept

3. For each "doesn't read right" point Gemini found:

   • Is this a real gap, or did Gemini miss something obvious?
   • If real: what shared context were you relying on that isn't in the document?

Log findings (append to decision-log.md):

### Blind Review: [document name]
**Date:** [auto]
**Gemini's understanding:** "[one-sentence summary]"
**Intended meaning:** "[what we actually meant]"
**Divergence level:** None / Minor / Major
**Gaps found:**
1. [gap]: Real gap / False positive — Action: [fix / ignore]
2. [gap]: Real gap / False positive — Action: [fix / ignore]
**Root cause:** [expression problem / direction problem / none]

Note: This is an AI-assisted review record, not a human decision entry. Mark it clearly as such in the log.

Step 3: Informed re-review (optional)

If Step 2 found major divergence and the document was revised, generate a prompt for a Gemini session that does have the original context:

The document below was revised after external feedback. You previously
understood my original intent for this project. Please evaluate:
1. Does the revision still align with my core goal?
2. Did clarity improvements sacrifice anything important?
3. Remaining issues?

Revised document:
---
[paste revised document]
---

Step 4: Devil's advocate (conditional)

Trigger condition: All reviewers (Claude + Gemini) agree the document is solid, with no substantive disagreements.

Complexity gate: Only trigger on documents involving core architecture decisions, strategic direction, or significant resource commitment. Skip for UI tweaks, copy changes, or well-scoped small features.

When triggered:

You are the most skeptical critic in [this domain]. Your professional habit
is finding fatal flaws that everyone else missed. You are especially suspicious
when "everyone agrees it's fine."

This document passed two independent reviews with no major issues.
That is precisely when you pay closest attention.

1. What 2+ assumptions in this document will be proven wrong within 6 months?
2. What paths did reviewers miss because they were anchored to the proposal?
3. If you were a competitor, how would you defeat this approach?
4. "If we must proceed, at minimum we must validate ______ first."

Document:
---
[paste document]
---

If fatal flaw found: flag it and suggest returning to revision. If no fatal flaw: confidence level is high. Document is cleared.

Output files

File	Purpose
prompts/gemini-blind-review.md	Ready-to-paste blind review prompt
decision-log.md	Review findings (append, marked as AI-assisted review)