| name | paper-review |
| description | Use when the user wants reviewer feedback on a draft paper before submission, mentions paper review, reviewer simulation, PC review, or rebuttal preparation |
Paper Review
Core principle: Realistic review requires DIVERSE PERSONAS. Five subagents embody distinct reviewers (unique domain, expertise 1-5, taste) and review in parallel. Each supports interactive QA with persistent memory.
Step 1 MUST complete before Step 2. Step 2 MUST complete before Step 3.
Main agent MUST NOT review — only coordinate: profiles -> dispatch -> collect -> QA.
Process Flow
digraph paper_review {
rankdir=TB;
"Start" [shape=oval];
"Existing reviews?" [shape=diamond];
"Ask: re-review, QA, exit?" [shape=diamond];
"Step 1: Generate 5 profiles\n[reviewer-profiles.md]" [shape=box];
"User approves?" [shape=diamond];
"Step 2: Discover template\nfrom templates/, dispatch 5 subagents" [shape=box];
"All 5 done?" [shape=diamond];
"Step 3: QA with reviewer?\n[qa-protocol.md]" [shape=diamond];
"Load profile+history,\nrespond in character\n(use scripts/qa-entry)" [shape=box];
"More QA?" [shape=diamond];
"Done — offer aggregate\n(scripts/aggregate-report)" [shape=oval];
"Start" -> "Existing reviews?";
"Existing reviews?" -> "Ask: re-review, QA, exit?" [label="yes"];
"Existing reviews?" -> "Step 1: Generate 5 profiles\n[reviewer-profiles.md]" [label="no"];
"Ask: re-review, QA, exit?" -> "Step 1: Generate 5 profiles\n[reviewer-profiles.md]" [label="re-review"];
"Ask: re-review, QA, exit?" -> "Step 3: QA with reviewer?\n[qa-protocol.md]" [label="QA"];
"Ask: re-review, QA, exit?" -> "Done — offer aggregate\n(scripts/aggregate-report)" [label="exit"];
"Step 1: Generate 5 profiles\n[reviewer-profiles.md]" -> "User approves?";
"User approves?" -> "Step 1: Generate 5 profiles\n[reviewer-profiles.md]" [label="no"];
"User approves?" -> "Step 2: Discover template\nfrom templates/, dispatch 5 subagents" [label="yes"];
"Step 2: Discover template\nfrom templates/, dispatch 5 subagents" -> "All 5 done?";
"All 5 done?" -> "Step 2: Discover template\nfrom templates/, dispatch 5 subagents" [label="no"];
"All 5 done?" -> "Step 3: QA with reviewer?\n[qa-protocol.md]" [label="yes"];
"Step 3: QA with reviewer?\n[qa-protocol.md]" -> "Load profile+history,\nrespond in character\n(use scripts/qa-entry)" [label="yes"];
"Step 3: QA with reviewer?\n[qa-protocol.md]" -> "Done — offer aggregate\n(scripts/aggregate-report)" [label="no"];
"Load profile+history,\nrespond in character\n(use scripts/qa-entry)" -> "More QA?";
"More QA?" -> "Load profile+history,\nrespond in character\n(use scripts/qa-entry)" [label="yes"];
"More QA?" -> "Done — offer aggregate\n(scripts/aggregate-report)" [label="no"];
}
Step 1: Profiles
Generate 5 reviewer profiles. Reference: reviewer-profiles.md. Output: review/reviewers/profiles.md.
Before Step 2: all 4 domains covered, expertise 1-5 spread (min one 1-2, one 5), distinct tastes, user approved.
Step 2: Parallel Review
List templates/ recursively to discover available review templates. Match by target venue. Load the matching review-guid.md.
Dispatch 5 subagents simultaneously. Each receives: profile + template + full paper + output path (review/reviews/reviewer<N>/review.md).
Task prompt: "You ARE Reviewer N. Use template at . Reflect your domain, expertise, and taste. Include ALL template sections. Do NOT write as generic AI."
Before Step 3: all 5 reviews complete, each follows template structure, each persona-consistent.
Step 3: Interactive QA
Reference: qa-protocol.md. Use scripts/qa-entry.cmd (Win) or scripts/qa-entry.sh (Linux) to append entries. Load profile + full QA history before every response. Answer in character.
Review MUST exist before QA. Append with script, not manually. Record Internal Reasoning for consistency.
Completion
Offer scripts/aggregate-report.cmd (Win) or scripts/aggregate-report.sh (Linux). Output: review/aggregate-report.md.
Guardrails
| Red flag | Reality |
|---|
| "I'll review without profiles" | No profiles = generic reviews. Step 1 first. |
| "All reviewers should be experts" | Non-experts reveal exposition gaps. Spread 1-5. |
| "I'll summarize all 5 into one" | Collapsing personas = missing the point. |
| "Same review criteria for all papers" | List templates/, discover types, match by venue. |
| "I remember the profile, no need to reload" | Load profile + QA history before EVERY response. |
| "Reviews look similar = consensus" | Similar reviews = persona collapse, not validation. |
