research-review

name	research-review
description	Get a deep critical review of research via Agent self-review (Opus critic role). Use when user says "review my research", "help me review", "get external review", or wants critical feedback on research ideas, papers, or experimental results.
argument-hint	["topic-or-scope"]
allowed-tools	Bash(*), Read, Grep, Glob, Write, Edit, Agent

Research Review via Agent Self-Review (Opus Critic Role)

Get a multi-round critical review of research work using a Claude Opus Agent sub-agent as an adversarial critic.

Constants

• REVIEWER_MODEL = claude-sonnet-4-6 — Self-review via Agent sub-agent (critic role)

Context: $ARGUMENTS

Prerequisites

No external MCP required. Reviews use the Agent tool to spawn a sub-agent acting as an adversarial ML reviewer.

Workflow

Step 1: Gather Research Context

Before calling the reviewer, compile a comprehensive briefing:

1. Read project narrative documents (e.g., STORY.md, README.md, paper drafts)
2. Read any memory/notes files for key findings and experiment history
3. Identify: core claims, methodology, key results, known weaknesses

Step 2: Initial Review (Round 1)

Use the Agent tool with a critic role prompt:

Agent prompt:
  You are playing the role of a rigorous external ML reviewer (NeurIPS/ICML level).
  Your job is to find weaknesses, not to be supportive. Be adversarial and honest.

  Research context:
  [Full research context + specific questions]

  Please identify:
  1. Logical gaps or unjustified claims
  2. Missing experiments that would strengthen the story
  3. Narrative weaknesses
  4. Whether the contribution is sufficient for a top venue

  Be brutally honest. Save your review to RESEARCH_REVIEW.md.

Save the Agent's response verbatim to RESEARCH_REVIEW.md.

Step 3: Iterative Dialogue (Rounds 2-N)

Each round is a new Agent call. Include the previous review as context in the prompt:

Agent prompt:
  You are continuing a multi-round review. Your previous assessment was:
  ---
  [Paste previous review from RESEARCH_REVIEW.md]
  ---

  The researcher has responded:
  [Your response to criticisms / new evidence / counterarguments]

  Follow-up questions:
  [Targeted follow-ups on actionable points]

  Please update your assessment and respond to the above.

Key follow-up patterns:

• "If we reframe X as Y, does that change your assessment?"
• "What's the minimum experiment to satisfy concern Z?"
• "Please design the minimal additional experiment package (highest acceptance lift per GPU week)"
• "Please write a mock NeurIPS/ICML review with scores"
• "Give me a results-to-claims matrix for possible experimental outcomes"

Step 4: Convergence

Stop iterating when:

• Both sides agree on the core claims and their evidence requirements
• A concrete experiment plan is established
• The narrative structure is settled

Step 5: Document Everything

Save conclusions to RESEARCH_REVIEW.md:

• Round-by-round summary of criticisms and responses
• Final consensus on claims, narrative, and experiments
• Claims matrix (what claims are allowed under each possible outcome)
• Prioritized TODO list with estimated compute costs
• Paper outline if discussed

Update project memory/notes with key review conclusions.

Key Rules

• Send comprehensive context in Round 1 — the sub-agent cannot read your files automatically
• Be honest about weaknesses — hiding them leads to worse feedback
• Push back on criticisms you disagree with, but accept valid ones
• Focus on ACTIONABLE feedback — "what experiment would fix this?"
• The review document should be self-contained (readable without the conversation)
• Include previous review text in each subsequent round's Agent prompt for consistency

Prompt Templates

For initial review:

"You are playing a senior ML reviewer (NeurIPS/ICML level). I'm going to present a complete ML research project for your critical review..."

For experiment design:

"Please design the minimal additional experiment package that gives the highest acceptance lift per GPU week. Our compute: [describe]. Be very specific about configurations."

For paper structure:

"Please turn this into a concrete paper outline with section-by-section claims and figure plan."

For claims matrix:

"Please give me a results-to-claims matrix: what claim is allowed under each possible outcome of experiments X and Y?"

For mock review:

"Please write a mock NeurIPS review with: Summary, Strengths, Weaknesses, Questions for Authors, Score, Confidence, and What Would Move Toward Accept."