alterlab-ieu/alterlab-academic-skills

📋 Table of contents

🧬 What is this?

AlterLab Academic Skills is a library of Agent Skills for academic work: literature review, manuscript writing and review, citation checking, research design and statistics, scientific databases and toolkits, and the Turkish academic system (YÖK, ÜAK, DergiPark, TÜBİTAK). Each skill is a SKILL.md instruction file, usually with reference documents and small helper scripts, that gives Claude a method for one kind of task.

Every skill does what it claims with no external API key, account or login: Claude’s own tools, keyless public APIs and local libraries. A credential may only add something (your own private account, a higher rate limit, a paid service you ask for by name), and each one is declared in the skill’s metadata.credentials. scripts/check_keyless.py enforces this in CI.

[!IMPORTANT] Treat what a skill produces the way you would treat a capable research assistant’s draft: useful, and something you check. The skills ask Claude to cite sources and to disclose AI assistance, but they do not make its output correct, complete, or ready to submit.

Built for faculty and researchers. Student skills live in the sister project AlterLab-FC-Skills.

🚀 Quick start

Option 1 — Claude Code plugins (recommended)

Validated with claude plugin validate:

/plugin marketplace add AlterLab-IEU/AlterLab-Academic-Skills
/plugin install alterlab-essentials@alterlab-academic-skills   # starter set, or any single domain:
/plugin install alterlab-turkish-academia@alterlab-academic-skills
/reload-plugins

Domain plugins: alterlab-core, -databases, -bioinformatics, -cheminformatics, -clinical-research, -data-science, -visualization, -writing-tools, -lab-integrations, -domain-specific, -document-tools, -research-tools, -finance-economics, -turkish-academia, -faculty-life, -methodology, -social-science-workflow, -workflows; bundles alterlab-essentials and alterlab-complete.

Option 2 — claude.ai / the Claude app

Upload one skill at a time: download skill--.zip from Releases and use Customize ▸ Skills ▸ Upload a skill (code execution must be on). These archives carry the skill’s shared files and license notices and drop frontmatter the uploader is reported to reject; skills that need a companion skill say so in the release notes. Packaging is tested here; acceptance by the live uploader is not. The alterlab--.zip archives are not uploads — unzip them into a skills directory.

Option 3 — Plain install or another agent

Into the Claude Code skills directory or the cross-tool ~/.agents/skills/ (needs python3). Skills a selected skill depends on are installed with it (--no-deps to skip):

git clone https://github.com/AlterLab-IEU/AlterLab-Academic-Skills.git && cd AlterLab-Academic-Skills
scripts/install.sh --list
scripts/install.sh core/alterlab-citation-verifier      # one skill (+ its dependencies)
scripts/install.sh turkish-academia                     # a domain

The skills follow the open Agent Skills format; docs/portability.md lists what carries over to other agents and what is Claude-Code-specific. Other runtimes have not been tested here.

🧭 Where do I start?

You do not need skill names: describe the task, and alterlab-skill-finder picks one and says which and why. If you want a starting point anyway:

Every skill, by domain: docs/CATALOG.md.

💡 Using it

Describe the task, or say “use AlterLab skills”, and alterlab-skill-finder picks a skill and says which and why. In Claude Code, the alterlab-workflows plugin adds seven multi-agent workflows:

Every workflow returns status: complete | partial | blocked with a coverage account (requested, completed, failed, unresolved) and returns a report path only after the report file has been read back. Separate agents of the same model avoid anchoring on each other but are not independent verification, and the reports say so.

✅ What has been checked, and what has not

📚 Domains

Every skill, by domain: docs/CATALOG.md · searchable: alterlab-ieu.github.io/AlterLab-Academic-Skills · machine-readable: skills.json.

🔎 Integrity tools: what they establish

• alterlab-citation-verifier checks that references exist and that their metadata agrees with Crossref, Semantic Scholar, Europe PMC, arXiv (Open Library for books), and doi.org, and flags retractions. Existence and metadata are reported separately: a resolving DOI does not clear an author the record does not list (PAC, with the unsupported names). A reference absent from every index that answered is NOT_FOUND — a prompt for a manual check, not a finding of fabrication — and an unregistered DOI is ID_INVALID (the identifier is wrong; the work is undetermined). Retraction status is flagged, not_flagged_in_checked_sources or unknown (Crossref is the only retraction source checked). An entry it could not check is unverified, never a pass; input it could not parse makes the run PARTIAL (exit 3), never a pass.
• claim_faithfulness.py compares a claim with the cited work’s abstract only. It never reports a claim as verified: a heuristic match is LEXICAL_MATCH_REQUIRES_REVIEW, a model judgment is MODEL_JUDGED_SUPPORT_REQUIRES_REVIEW, and guards make it abstain on wrong figures, reversed comparisons, population mismatches, and tentative evidence. Read the cited passage before treating a claim as supported. See docs/integrity.md.

🔬 Check it yourself

uv run --frozen pytest tests/ -q                       # the validation suite
uv run python scripts/run_evals.py --strict            # eval-file structure (static)
uv run python scripts/run_evals.py --behavioral --skill alterlab-docentlik-eligibility
uv run python scripts/run_benchmark.py                 # the five-task benchmark

The live commands use the claude CLI’s own login, so they run inside an authenticated Claude Code environment; each run writes JSON results to eval-results/. A behavioral run exits 0 only if every selected case produced a verdict and passed. See docs/evals.md.

🔗 Sister projects

📜 Licensing

The repository is MIT-licensed, but each skill’s content is under the license in its own license: frontmatter field, and those differ: 201 MIT, 19 Apache-2.0, 4 CC-BY-NC-4.0, 4 GPL-3.0, 2 GPL-2.0, 1 LGPL-3.0, 1 CeCILL-2.1, 2 BSD-3-Clause, 2 CC-BY-4.0, 1 CC-BY-3.0, 4 CC0-1.0. The four core pipeline skills (alterlab-deep-research, alterlab-paper-writer, alterlab-paper-reviewer, alterlab-research-pipeline) are non-commercial (CC-BY-NC-4.0, from their upstream). Check a skill’s license before redistributing it or using it commercially. Every installed or packaged skill carries LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md; lineage is in PROVENANCE.md and THIRD_PARTY_NOTICES.md.

📝 Cite

@software{alterlab_academic_skills_2026,
  title   = {AlterLab Academic Skills},
  author  = {{AlterLab Creative Technologies Laboratory, Izmir University of Economics (IEU)}},
  year    = {2026},
  version = {4.1.0},
  url     = {https://github.com/AlterLab-IEU/AlterLab-Academic-Skills}
}

CITATION.cff has the full metadata. Release history is in CHANGELOG.md.

🤝 Contributing and credits

Contributions are welcome — see CONTRIBUTING.md. Behavioral evidence (new benchmark cases with anchored answers, or a recorded run of a skill’s evals) is the most useful contribution right now.

git checkout -b improve/alterlab-citation-verifier
# edit skills/core/alterlab-citation-verifier/SKILL.md, then run the checks above
git commit -m "improve: alterlab-citation-verifier — "

The suite began as a content fork of K-Dense’s scientific-agent-skills (MIT) and adapts the core research pipeline from Cheng-I Wu’s academic-research-skills (CC-BY-NC 4.0).