Skip to content

kalarislabs/research-agent-skills

v1.2.0MIT

283 Agent Skills for researchers: scientific and research paper writing, journal formats, literature review, citations, data science, ML research and domain science.

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning.

Fixed vulnerabilities are listed under Security in the release that contains the fix, with the advisory or CVE identifier. Security advisories are also published at https://github.com/KalarisLabs/research-agent-skills/security/advisories.

[Unreleased]

[1.2.0] - 2026-11-16

This release makes the skills do what their descriptions say. Every claim a SKILL.md makes about a script is now either true in the code or caught by CI, and the library grows to 283 skills.

Fixed

  • citation-verification now detects retracted papers: Crossref update notices (retraction, withdrawal, removal, expression of concern, with Retraction Watch as source) and OpenAlex is_retracted. New status retracted; an expression of concern reports uncertain. Dataset and software DOIs resolve through DataCite instead of failing as doi-not-found.
  • verify_citations.py degrades per source: an unreachable Crossref, DataCite or OpenAlex gives uncertain, never a false not-found. HTTP 429 honours Retry-After. --help no longer crashes on the --annotate text.
  • verify_citations.py --annotate writes % UNVERIFIED comments above flagged entries, as the skill promised.
  • systematic-review-prisma refuses counts that have no logged search (search_log: source, platform, exact query, date, hits, limits; PRISMA 2020 item 7 and PRISMA-S). --allow-unlogged is for legacy counts only. --search-table writes the supplement table.
  • Retrieved abstracts and paper text are treated as untrusted data in ten search and review skills, so an instruction embedded in a paper is quoted, not followed.

Added

  • New skill paper-card: one verified card per paper. check_card.py confirms every claim's page, verbatim quote and numbers against the converted text, evaluates derived numbers safely, and builds an independent second-reviewer packet. A claim is accepted only when both checks pass.
  • New skill fulltext-acquisition: fulltext_status.py finds which included records lack full text and writes missing.csv and retrieval.md (offline by default; --resolve-oa uses Unpaywall); match_pdf.py identifies downloaded PDFs by the DOI, PMID or title inside them and attaches only text-verified matches.
  • paper-corpus-rag rebuilt around locate-then-read: paper_convert.py (page markers, quality grade per page, symbol-font repairs logged in a sidecar), heading-bound page-aware chunks, hybrid retrieval (FTS5 BM25 plus optional embeddings fused by reciprocal rank), read and eval (recall@k, MRR) subcommands. Existing indexes need index --rebuild.
  • unslop-academic-writing: slop_check.py --claims lists what each paragraph claims and whether it carries evidence, and the skill asks for that before any rewriting.
  • tools/claim_audit.py (CI and make claims): fails when a SKILL.md names a script, flag, host, service or status its scripts do not have. tools/check_skill_count.py fails when prose quotes a stale skill count.
  • tools/validate.py enforces an eval floor for script-backed original skills (10 prompts, 3 near-misses); six older skills are grandfathered as warnings in tools/eval_floor_baseline.txt.
  • Citation benchmark: retracted entries (Wakefield 1998, Séralini 2012) and a Zenodo dataset; the run gates on retraction_recall.
  • unpaywall.org and europepmc.org in security/allowed-domains.txt.

Changed

  • Trigger-routing evals: prompts that quote a skill's own scripts or errors are marked follow_up: true and skipped by the BM25 router. The reviewed suite doubled to 106 prompts, so the PR-gate numbers are harder (hit@1 0.90, hit@3 0.99, false-trigger 0.08, was 1.00/1.00/0.12 on 53 prompts).
  • Benchmark README tables, charts and skill-count claims regenerated for 283 skills.

Changed (earlier in this cycle)

  • README repositioned around turning any agent into a research co-scientist: a research-stage map of skills and a "Works in any harness" section covering skills.sh, the plugin marketplace and manual installs.
  • Distribution is skills.sh, the Claude Code plugin marketplace, GitHub CLI and the checksum-verified install.sh/install.ps1 installers. The research-agent-skills CLI is not published to npm; README, docs, generated skill pages and metadata no longer point to npm.

Added

  • Eval suites for all 283 skills (evals/<skill>/evals.json): realistic prompts plus near misses naming the better skill. Suites for adapted skills were generated from each SKILL.md and are marked pending review.
  • Import-time enrichment of adapted skills (tools/enrich.py): oversized skills split into references/, dead links repaired, category-specific agent operating procedures (environment check, small first run, validation, failure-recovery table, integrity rules) and computed related-skill handoffs.
  • Trigger-rich description overrides for 84 adapted skills (third_party/description-overrides.yaml, generated by tools/generate_skill_data.py and validated).
  • Offline smoke tests for every skill script: all 497 compile; 327 standard-library CLIs run --help on every CI OS (tests/test_skill_scripts_smoke.py).
  • tools/trigger_bench.py --sample N --seed S for reproducible subsamples with the paid Claude router.
  • Search tags for all 283 skills. Tags for adapted skills live in third_party/tag-overrides.yaml and are applied at import; tools/validate.py warns when a skill has none.
  • llms-full.txt: every skill with its full description, tags and install command.
  • Generated skill pages now carry tags, example requests from the eval suites, related skills, field-guide links, every install route and a sentence-complete meta description.
  • Marketplace plugins declare version, author, homepage, repository, license, category and keywords (the most common tags of their skills).
  • .zenodo.json and codemeta.json for DOI minting and research-software indexing.
  • 1280×640 GitHub social preview image (PUBLIC/social-preview.png) and a 59 KB WebP README banner.

Changed

  • Marketing residue (GitHub star counts, user numbers) is stripped from imported skills.
  • The 500-line SKILL.md budget is now enforced for every skill (the size baseline is empty).
  • Adapted-skill quality rose from a mean of 80.9 to 95.2. Claude-router trigger accuracy on a 100-prompt sample of all evals is hit@3 0.96 with no false triggers.

Fixed

  • openpiv/run_example.py explains a missing dependency instead of raising a traceback.
  • The release workflow's npm job no longer skips when the optional vendor security gate is skipped.
  • The npm installer is no longer described as unreleased.
  • plugin.json, the npm package, CITATION.cff and SUPPORT.md point at one documentation URL.

Removed

  • Unused MkDocs configuration (mkdocs.yml, docs/overrides/) and the docs dependency group. The site is built by Mintlify, which never loaded the MkDocs Open Graph and JSON-LD templates.
  • MkDocs-only ??? admonitions in the FAQ, now plain headings that Mintlify renders.

[1.1.0] - 2026-10-01

Added

  • unslop-academic-writing: removes AI slop from research prose (stock vocabulary, empty emphasis, hedge stacks, formulaic signposting, monotone rhythm) with a zero-dependency linter for Markdown, LaTeX, text and Word files.
  • Skill verification tooling: static quality rubric (tools/skill_quality.py), trigger-routing benchmark (tools/trigger_bench.py, BM25 and Claude routers), with-vs-without-skill task benchmark with a blind judge (benchmarks/task_evals), labeled citation-verification benchmark and a human-vs-model slop benchmark.
  • benchmarks.yml workflow: quality and routing gates on every PR, weekly citation benchmark, manual paid benchmarks.
  • Documentation site restructure: getting started, guides (papers, theses, no-slop writing, systematic reviews, skills by field), reference (CLI, architecture, benchmarks, security, contributing), FAQ, JSON-LD metadata.
  • Generated llms.txt for AI search engines and agents.
  • .editorconfig, .pre-commit-config.yaml, Makefile, SUPPORT.md, GOVERNANCE.md.

Changed

  • Repository moved to KalarisLabs/research-agent-skills; code owner @saynchowdhury.
  • Journal-format and other original skills gained numbered workflows and explicit integrity rules.
  • Sharper descriptions for cover-letter-to-editor and rebuttal-and-response-to-reviewers to prevent mis-triggering.

[1.0.0] - 2026-10-01

Added

  • 278 agent skills across 19 categories: research writing, journal formats, literature review, ideation and design, data science, visualization, knowledge graphs and RAG, scientific databases, life sciences, chemistry, clinical research, physical sciences, lab automation, research automation and ML research engineering.
  • 21 original skills, including journal formats (Nature Portfolio, Science, Cell Press, IEEE, ACM, Elsevier, Springer LNCS, PLOS, APA 7, arXiv), citation verification, BibTeX hygiene, systematic reviews (PRISMA 2020), reference-manager interop, paper-corpus RAG, research knowledge graphs, abstracts and titles, reviewer responses, cover letters, reproducibility statements and a research skill creator.
  • research-agent-skills CLI: one-command install for Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode, Windsurf and .agents/skills harnesses, with checksum verification, integrity checks (doctor), update and uninstall.
  • Standalone install.sh / install.ps1 installers with SHA-256 verification.
  • Security CI: spec validation, prompt-injection lint, Cisco AI Defense skill-scanner, CodeQL, Semgrep, Bandit, TruffleHog, OSV/pip-audit/npm audit, zizmor, actionlint, OpenSSF Scorecard, signed releases with build provenance.

Fixed

  • Windows console crashes (cp1252) in skill scripts that print non-ASCII text.
  • experimental-design: fractional allocation ratios no longer silently drop an arm.
  • analytical-method-validation: ragged CSV rows exit cleanly with code 2; JSON and CSV inputs are normalized identically.
  • arbor: set-evidence validates status; node lists sort n1..n10 numerically.
  • deeptools: BAM index detection no longer breaks when a parent directory contains .bam.
  • Doubled uv uv pip install commands.