| name | scientific-review |
| description | Hierarchical review of scientific documents (papers, grant proposals) without editing them. Layer 1 spawns naive skim and adversarial critics that file findings as GitHub issues; Layer 2 triages each finding simplification-first (cut before patch). Runs coarse to fine — document skeleton, sections, paragraphs — locking user decisions between stages, then a sentence-level style check. Use for any review or critique request on paper or proposal prose. For writing, rewriting, or wording edits, use the scientific-writing skill instead. |
Scientific Review
Two-layer staged review. Layer 1 finds problems; Layer 2 picks the cheapest sound fix, cutting logic over adding it. Findings live in GitHub issues, not chat. This skill decides what to change; edits run under scientific-writing.
workflow.md — stages, layers, issue conventions, decision points. Read first on any review request.
agents/skim-critic.md — L1: naive busy reader, reads back the skim extract.
agents/adversarial-critic.md — L1: hostile expert, attacks claims and evidence at the stage's scope.
agents/simplifier.md — L2: triages issues (patch/cut/reject), owns the risk–return audit.
agents/style-critic.md — final pass: sentences vs scientific-writing's style.md.
Each agent file fully defines one subagent: role, inputs, output. Spawn with the file's content + only the inputs it names.
Principles
- Critics are naive. Only the manuscript (skim critic: only the extract) + repo/labels for filing. No history, no skill files, no takeaway list. Naivety is the measurement instrument.
- Simple beats complex. A reader who cannot follow a braided defense rejects. If a patch adds a qualifier chain or a new thread, cut the element and move its function to a stronger instrument already in the document.
- Findings are issues. Agents return issue numbers + one-liners only; coordinator triages from
gh issue list, opens bodies selectively.
- Decisions lock. User rulings recorded on the issue; later stages never reopen them.
Models
Critics + simplifier: Opus. Style critic: Haiku. Post-review edits (scientific-writing): Sonnet, Haiku for wording.
Reporting
Japanese. Per stage: findings by severity with L2's recommendation; needs-decision issues as explicit questions. Never translate the document's own prose.