Skip to content

graphe:paper-review

A pre-submission novelty and reviewer pass for a draft paper. It checks whether each contribution is actually new, what related work the draft misses, whether its claims hold up, and whether it overlaps a lab or co-author line that must be disclosed. Every judgment is grounded in a real retrieved paper, never asserted from memory.

Part of graphe: /plugin install graphe@techne, then /graphe:setup.

When to use

  • A first-author draft is approaching submission: "review <name> for novelty", "what related work am I missing", "is this contribution new".
  • NOT for copy-editing prose or for a paper with no stated contributions yet; scaffold first with graphe:paper.

Usage

/graphe:paper-review <name | path/to/doc.tex> [--from <repo>]

Reads the ## paper-review section of .claude/skill-context.md (lab_line, the overlap source of truth; retrieval; mailto, for OpenAlex's polite pool and also read from OPENALEX_MAILTO; and report, the report path), extracts the draft's contributions, retrieves comparable prior work from OpenAlex (web-search backstop; Semantic Scholar only with a key saved in /plugin configure graphe@techne or exported in the shell; an OpenAlex key raises OpenAlex's budget). Only derived keyword queries leave the machine, never the paper text, and writes papers/<name>/novelty-review.md (or <stem>.novelty-review.md beside a .tex path): citation integrity, novelty per contribution, related-work gaps, claim-support, a lab-overlap disclosure section, and a provenance log of every query. Advisory only: it never edits the draft.

What it grounds

A mandatory pre-check (ยง0) first verifies the draft's own citations against authoritative records (arXiv API / OpenAlex / DOI): a fabricated id, an id that resolves to a different paper, or a mischaracterized claim about a cited work is a stop-ship finding. Then every novelty verdict cites a paper retrieved that run with a verbatim snippet, logged in a provenance appendix so the search is re-runnable. A verdict with no retrieved record is marked unverified. The lab-overlap section surfaces adjacency and a COPE disclosure checklist but renders no verdict, since the solo-vs-lab boundary is a human agreement.

Claim-support also runs the Claims checks in _shared/plain-prose.md: numbers against the code or data behind them, method descriptions against the code, summaries of cited papers against their full text where open access, and a proposal's promises against the time it has. A manuscript missing the sections its venue expects, such as Related Work or References, is a stop-ship finding.

The pass also applies _shared/paper-quality.md: the every-sentence test, a comparison against the venue conventions sheet, an audit of CUTS.md against the call's required elements, integrity guards (ethics or IRB claims, restricted data, "not in a published figure" promises), and registered-report tense. Reviewer lenses find weaknesses; a finding counts only if fixing it makes the paper clearer, more rigorous or more concise.

See also

Full instructions

SKILL.md on GitHub holds the full instructions Claude follows for this skill.