Review Skill
Stacking is the default failure mode of autonomous review projects: "Smith et al. did X, found Y. Jones et al. extended to Z..." paragraph after paragraph, with no unifying thesis and no synthesis of tensions. Top-tier reviews do the opposite — they lead with a claim, pull disparate work into one frame, and leave the reader smarter than when they started.
This skill encodes the rhetorical and structural conventions of landmark reviews across 10 domains, mined from real corpus text (RMP, Chem. Rev., Nat. Rev. X, Annu. Rev., Bull. AMS, Science / Nature Reviews, and the domain-specific tier-1 venues). Inject a domain's style guide, follow the 3-step pipeline, obey the hard rules — and the output should pass the Turing test for an editor at its target venue.
When to use which venue voice
┌──────────────────────────────────────────────────┬──────────────────┐ │ Target length / audience │ Voice │ ├──────────────────────────────────────────────────┼──────────────────┤ │ Monograph (50–150 pp), pedagogical, archival │ RMP / Chem.Rev. │ │ → equations, boxes, glossary, deep derivation │ / Phys.Rep. │ │ Specialist (20–40 pp), claim-dense, tutorial │ Annual Reviews │ │ → thesis-per-section, figure-rich │ │ │ Short assessment (5–15 pp), stance-forward │ Nat. Rev. X / │ │ → abstract-caliber thesis per ¶, citation-dense │ Trends / Curr.Op. │ │ Opinionated essay (5–20 pp), single author │ Nat Comment / │ │ → warm first-plural, polemical beat │ Nobel lecture │ └──────────────────────────────────────────────────┴──────────────────┘
See references/decision_tree.md for full venue taxonomy.
The 3-step pipeline (mandatory)
Skip no step. Each exists because stacker projects skip exactly this step.
Step 1 — Outline with thesis per section (BEFORE any prose)
Produce notes/report_outline.md first (canonical path — the finish-gate's
outline check reads exactly this file; first line MUST be type: survey).
For the annotated gold-standard skeleton, see
skills/review/references/exemplar_survey_outline.md:
markdowntype: survey # <title> ## §1 Introduction **Thesis:** <one sentence. What is this review claiming about the field?> **Unifying framework:** <concept / equation / taxonomy that ties everything> **Fold in:** Paper A (framework), Paper B (canonical result), … ## §2 <topic> **Thesis:** <claim this section argues, not a topic label> **Fold in:** Paper C, D, E **Synthesis move:** <comparative_table | contrast_pair | unifying_equation | …> …
Gate: do not draft prose until a PI / reviewer / user has signed off on this outline. This is the single most effective anti-stacking move — if you cannot articulate a thesis per section in one sentence, the section will stack.
Step 2 — Draft, reading domain style guide before each section
Before drafting §N:
- Read
style_guides/<domain>.md— the entire guide, every time, fresh. - Read §N's outline entry.
- Write prose that obeys the guide's claim-leading, synthesis-move, and transition conventions. Quote primary papers only where a claim needs evidence, not to fill the paragraph.
Step 3 — Synthesis rewrite pass (POST-draft)
After the full draft:
- For every paragraph, check its first sentence. If it starts with "Smith et al." / "Jones and coworkers" / "In [42]" / any author name → rewrite it to lead with a claim about the phenomenon, and push the citation to mid-paragraph or parenthetical.
- For every section, check its first paragraph. If it starts with orientation ("This section discusses…") instead of thesis → rewrite the first sentence to state the claim.
- For every section-to-section boundary, add a transition sentence that
names the bridge (cause→effect / contrast / zoom-in / caveat /
historical pivot). See
references/transition_moves.md.
Hard rules
These are non-negotiable. Violations should trigger rewrite before output.
- No paragraph begins with an authorship phrase. "Smith et al." / "Jones and coworkers" / "In [42]" as paragraph-openers are banned. Lead with a claim about the phenomenon; attribute parenthetically or mid-paragraph.
- No section begins with "This section…" / "We now discuss…" / "Here we…" Lead with a thesis sentence about the field.
- No uncited claim, no un-claiming citation. Every citation must support a specific claim. Every claim beyond common knowledge gets a citation. Drive-by citations ("for a review see [7, 8, 9]" without claim anchoring) are a stacker tell — use them sparingly and only at section boundaries.
- Fold, don't list. When multiple papers make compatible findings, synthesize them into one claim with a shared citation cluster. When they disagree, make the tension the subject of the paragraph.
- Use domain voice. If writing physics, obey
style_guides/physics.mdon citation style (narrative vs parenthetical), equation density, section title voice. Do not import chemistry's taxonomy-monograph habits into a physics RMP draft.
Style guides (one per domain)
Auto-loaded by the agent when <domain> is set. Each is ~1,000–2,000 words of
prose distilled from 8–12 landmark reviews in that domain.
| Domain | Venues mined |
|---|---|
physics | RMP, Phys. Rep., Nat. Rev. Phys., Annu. Rev. Cond. Matt., Quantum, Science |
chemistry | Chem. Rev., Chem. Soc. Rev., Nat. Rev. Chem., Science |
biology | Cancer Discov., Nat. Rev. Genet., Science, eLife, Nat. Struct. Mol. Biol. |
medicine | NEJM, Nat. Med., Nat. Rev. Drug Discov., Nat. Rev. Genet. |
mathematics | SIAM Review, Acta Numerica, Bull. AMS, Probab. Surv., Curr. Dev. Math. |
computer_science | Nature, Science, NEJM, IEEE SPM, Nat. Mach. Intell., Annu. Rev. Fluid Mech., Comm. ACM |
earth_environment | Science, Rev. Geophys., Nature, Nat. Rev. Earth Env., PNAS, ESSD |
astronomy | ARA&A, Nat. Rev. Phys., A&A, ApJ Letters, PRX, CQG |
economics | Annu. Rev. Econ., JEP, AER, Handbook of Macroeconomics |
materials | Nature, Nat. Rev. Mater., Science, npj Comput. Mater., Nat. Energy, Chem. Soc. Rev. |
References
references/anti_patterns.md— the stacking tells (paragraph openers, drive-by citations, "zoo of results" paragraphs, untriaged citation clusters). What to not write.references/synthesis_rubric.md— the 7 canonical moves for unifying disparate work (unifying equation, comparative table, taxonomy tree, contrast pair, timeline, platform comparison, phase diagram).references/transition_moves.md— grammar of section-to-section and paragraph-to-paragraph movement.references/decision_tree.md— venue selection by length + audience + voice.
First-time usage
In a review project:
bash# 1. Copy the relevant style guide into your project (optional; skill reads # from here by default) cp skills/review/style_guides/physics.md report/style_guide.md # 2. Draft outline.md FIRST per Step 1 above # 3. Have PI review the outline — this is the gate # 4. Draft prose per section, reading the style guide before each section # 5. Synthesis rewrite pass per Step 3 # 6. Compile
Syncing style guides from source
Style guides are mined from real corpus text in the sister repo
review_style_skills. To refresh:
bashbash skills/review/scripts/sync_style_guides.sh
This re-copies data/style_guides/*.md from ~/Documents/review_style_skills
into skills/review/style_guides/.
