review · diff
git:20260816.d958232 to git:20260920.ce7b9ba
37 added, 12 removed. Audit A to A.
---
name: review
description: Review a manuscript or chapter as an external reviewer, or run an own-work self-review in a fresh-context clean room, producing anchored findings and a recommendation.
- allowed-tools: Read, Glob, Grep, Agent
+ allowed-tools: Read, Glob, Grep, Agent, Bash
---
# /review — Manuscript Review
## Modes
- **external** (default): review another author's work as submitted.
- **own-work**: review the user's own draft. The clean room is real, not
declared: spawn ONE fresh-context subagent whose prompt contains only the
manuscript text (and explicitly listed evidence files), and relay its
findings. Never present a same-session read-through as a clean-room review —
the drafting conversation is context contamination by definition.
## Evidence boundary
Judge the manuscript **as submitted**. Do not use prior chat memory, unstated
author intentions, or model background knowledge as evidence for or against a
claim. If a claim cannot be assessed from the manuscript and its cited
sources, say so — that is itself a finding.
## Findings format
- Every finding is anchored (section/paragraph or quoted span) and typed:
+ Write findings to a TSV the author can check, not prose they must take on
+ trust. A declaration of what was read, then one row per finding:
- | # | Type | Anchor | Finding | Severity |
- |---|------|--------|---------|----------|
+ ```
+ # reviewed: chapters/ch2_background.md, chapters/ch3_conceptual_framework.md
+ location source problem
+ chapters/ch2_background.md:142 literature/reading_notes/Foo_NOTES.md The sentence attributes a claim the note does not contain.
+ ```
- Types: `claim-exceeds-evidence` | `gap-contribution-mismatch` | `method` |
- `structure` | `clarity` | `citation`. Split findings three ways: defects that
- block the recommendation, improvements that would strengthen it, and
- questions the author must answer. Consult `references/argument-checklist.md`
- for the interrogation checklist and examiner-attack pre-mortem when the
- review targets argument quality.
+ - **location** is `file:line`, never a section name or a quoted span. A section
+ name cannot be told apart from a plausible one; a line number resolves or it
+ does not.
+ - **source** is the archived original or evidence file the finding rests on, or
+ `-` when the finding is about the text itself (structure, clarity, an
+ internal contradiction).
+ - **problem** is one sentence. Type and severity go in that sentence if they
+ matter; a separate taxonomy column was never read.
+ **Run the checker before handing the report over:**
+
+ ```
+ python3 .claude/skills/review/scripts/audit-review-findings.py \
+ --base-dir <manuscript> --findings <findings.tsv>
+ ```
+
+ It resolves every anchor, rejects a finding about a file the review never
+ declared it read, and exits 2 when the report does not say what it read — a
+ review with no declaration cannot be told apart from one that examined nothing.
+ It does **not** judge whether a finding is right. Split the rows three ways in
+ the accompanying note: defects that block the recommendation, improvements that
+ would strengthen it, and questions the author must answer. Consult
+ `references/argument-checklist.md` (repository root) for the interrogation
+ checklist and examiner-attack pre-mortem when the review targets argument
+ quality.
+
## Recommendation vocabulary
Exactly one of: `accept` | `minor_revision` | `major_revision` |
`reject_resubmit` | `reject` | `no_recommendation`.
## Stop conditions
Stop and report a blocker instead of continuing if asked to: fabricate
reviewer consensus, citation support, or experimental findings; review text
that is not available as submitted; or convert this review into a rewrite —
/review never edits the manuscript.
## Constraints
1. Never rewrite or patch the manuscript; output findings only.
- 2. No emoji. British English.
- 3. Own-work mode without a subagent available: state plainly that the review
+ 2. Every finding resolves: run the checker and report its exit code alongside
+ the findings. An unresolved anchor is withdrawn, not explained.
+ 3. No emoji. British English.
+ 4. Own-work mode without a subagent available: state plainly that the review
is NOT clean-room and label the output accordingly.