claude-skills-research

Claude Code skills for supporting academic research in machine learning and computer science: reviewing papers, verifying bibliographies, finding related work, comparing a draft against prior art, and running experiment campaigns on HPC clusters.

These grew out of real reviewing work. Every design decision that looks odd is there because something failed on a real paper, and the reasoning is recorded in the skill itself rather than lost.

Install

git clone https://github.com/chgagne/claude-skills-research ~/.claude/skills

Or clone elsewhere and symlink the ones you want into ~/.claude/skills/. Claude Code discovers a skill by the frontmatter in its SKILL.md.

No dependencies. Every tool is Python 3.10+ standard library only — no pip install, no virtualenv. A guard test in each skill fails the suite if a third-party import creeps in.

Optional configuration

Setting	Effect	How
SCHOLARLY_MAILTO	Puts Crossref/OpenAlex requests in their polite pool — faster and more reliable. Optional; everything works without it.	export SCHOLARLY_MAILTO="[email protected]" or write it to ~/.config/scholarly/mailto
S2_API_KEY	A dedicated 1 req/s lane on Semantic Scholar instead of a contended anonymous pool.	Request one free, then export S2_API_KEY=… or ~/.config/scholarly/s2_key
pdftotext	Lets comparing-papers read PDFs when LaTeX source is unavailable.	brew install poppler / apt install poppler-utils
sympy	Lets verifying-proofs confirm an algebraic identity rather than merely failing to refute it. Absent, those steps report UNVERIFIED and the run exits 2.	pip install sympy
z3	An opt-in escape hatch in verifying-proofs for a concrete inequality over reals or integers. Not used by any default run.	pip install z3-solver
latexmk	Builds the per-theorem PDFs in explaining-derivations. Absent, the .tex is still written and the run exits 2.	any TeX distribution

The proof skills never install any of these. A missing checker is a question for you, and the run degrades — naming what it could not check — rather than failing or pretending.

API responses are cached under ~/.cache/scholarly/, so re-runs across drafts are close to free.

The skills

Paper review

reviewing-paper-sources — the orchestrator. An eight-phase review that treats a paper as an evidence-gathering exercise rather than a reading exercise: compile the sources yourself, render the figures, recompute the headline statistic, audit every reference, then try to refute your own findings before delivering them. Phase 0 asks which optional modules below to run. Handles both internal review (a colleague's draft, with proposed fixes) and external refereeing (diagnosis only).

verifying-bibliography — finds incorrect and fabricated references. Resolves a stable identifier first, then compares the record back against the .bib. That inversion is the whole point: searching by the title you are trying to verify is circular, and a fabricated title that accurately describes the right paper will retrieve it and pass. Reports author problems as set differences (+Bach, +Dragicevic, −Howe) rather than a similarity score, because a scalar hides exactly what matters. Considerable care goes into not crying wolf: BibTeX's and others, LaTeX ties inside surnames, Unicode dash variants, records held in another script, and institutions listed where a .bib lists people are all normalised away, because a checker that reports fabrication on and others teaches its user to ignore it.

surveying-literature — finds work a draft should have cited, and maps fields. The gap sweep is a search method the agent runs: eight web searches built on names it knows (methods, systems, authors), then every candidate confirmed for free against Crossref and arXiv by lookup-titles.py instead of being searched for again. A script that expands through the citation graph across five engines, and clusters a field, is optional. Measured on held-out arXiv drafts in _evals/: the gain over the bare agent on hard references was +6 to +18 points on earlier sets and +4.6 on the latest; the current version costs 0.75 of the previous.

comparing-papers — puts a draft and its closest prior work side by side on the axes reviewers attack: training scale, seeds, compute, protocol, metrics. Fetches LaTeX source in preference to PDFs, because the numbers that decide reviews live in appendices. Computes a note only where the relation is arithmetic — a scale ratio, a seed count — and elsewhere shows two quoted passages and stops, because "these protocols differ" is a judgement.

verifying-proofs — checks the mathematics rather than the claims. The design inversion is that it refuses to translate LaTeX into a computer algebra system: parse_latex needs antlr4 and its grammar has no expectations, no \operatorname{}, no norms and no user macros, so a translator's bugs would surface as false counterexamples — the one failure a proof checker cannot survive. Instead it writes one auditable check script per step, carrying the source LaTeX, each symbol's domain and where that domain came from, and the side conditions the step needs. A symbol whose domain the paper never stated can never produce a counterexample; on one real paper 54 of 61 symbols were in that position, and sampling them freely would have manufactured dozens of errors against correct mathematics. Its default mode needs nothing external at all and still finds inductions with no base case, claim dependency cycles, restatements that drop a hypothesis, and divisions by quantities nobody proved non-zero.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗