AI Research Skills
17 portable SKILL.md research skills for Claude Code, Codex, Cursor and other AI clients, covering an eight-stage lifecycle from literature discovery to manuscript release.
- Source repo
- WenyuChiou/ai-research-skills
- Stars
- ★ 389
- Last updated
- today
- License
- MIT
- Primary language
- Python
- FA score
- 77/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude CodeCodex · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A grad student who has read 50 papers but can't tell which gap is thesis-worthy: use gap-to-topic for a three-gate decision dossier.
- Not a fit if
- Users wanting a zero-config turnkey product; teams needing corpus-scale validation or vendor support; researchers outside water resources / agent-based modeling
- Source review
- 77/100 · Some gaps
What does this agent do, and when should you use it?
AI Research Skills is a portable SKILL.md catalog (registry) that structures research into eight stages from literature discovery to manuscript release, with 17 skills and 5 installable plugins. The core asset is the portable SKILL.md instruction file; the Claude Code marketplace is just a convenience layer, and Codex CLI, Cursor, Gemini CLI, Hermes, and generic API clients can all load the same SKILL.md. An optional runtime, research-hub (pip install research-hub-pipeline), provides executable, resumable workflows, while an optional agent-collab-harness adds policy/checkpoint governance. The catalog rests on three design principles: manifests (.research/, .paper/) that persist research context across AI sessions, YAML schemas with anti-leakage rules governing cross-skill handoffs, and character-driven routing that delegates mechanical work to Codex/Antigravity. Skills integrate with Zotero, Obsidian, and NotebookLM, all three of which are optional inside research-hub. Automation pauses at five explicit human gates; read-only and reversible steps proceed automatically.
Once installed, skills trigger from natural language: literature-triage-matrix builds a matrix comparing papers by method, data, claim, and limitation; gap-to-topic produces a topic_dossier.gaps.yml through three go/no-go gates (is the gap open? a real contribution? feasible?); research-design-helper turns the gap into design_brief.md via Socratic dialog; research-context-compressor writes .research/ manifests so future AI sessions skip the rescan; paper-memory-builder generates .paper/claims.yml with a status enum and anti-leakage rule—a claim with empty evidence_artifacts is forced to status: gap with a gap_reason. research-workflow-orchestrator is a resumable end-to-end controller coordinating orient→scope→discover→synthesize→design→execute→write→release. research-hub (a Python CLI) chains arXiv/Semantic Scholar search, Zotero ingest, Obsidian note sync, and NotebookLM brief verification. codex-delegate and antigravity-delegate hand Claude's work to Codex CLI / Antigravity CLI, with research-hub-multi-ai writing a coordination file each delegate reads instead of the parent context. Every stage's output is the next stage's structured input; handoffs are documented as YAML schemas, and downstream skills refuse malformed handoffs (e.g. status: gap claims without gap_reason are rejected at Stage 7).
- A grad student who has read 50 papers but can't tell which gap is thesis-worthy: use gap-to-topic for a three-gate decision dossier.
- A PhD candidate building a comparative literature review: use research-hub for search/ingest, then literature-triage-matrix for a method/data/limitation matrix.
- A manuscript author worried about AI-sounding prose and hallucinated citations: use paper-memory-builder's claims.yml anti-leakage rule plus academic-writing-skills for terminology, repetition, and overclaim audits.
- A multi-tool user designing with Claude, scaffolding code with Codex, and finishing mechanical work in Antigravity: use research-hub-multi-ai for character-driven routing with coordination files.
- Research support staff already running Zotero/Obsidian/NotebookLM: install zotero-skills and research-hub to unify them into one library-management loop.
- Agent/harness builders: reuse the architecture, state, HITL, and recovery design documented in docs/for-agent-harness-builders.md.
How do you install or deploy this agent?
Fastest path (Claude Code, ~30 seconds):
claude plugin marketplace add WenyuChiou/ai-research-skills
claude plugin install research-workspace@ai-research-skills
claude plugin listAdd optional plugins as needed:
claude plugin install academic-writing-skills@ai-research-skills
claude plugin install zotero-skills@ai-research-skills
claude plugin install codex-delegate@ai-research-skills
claude plugin install antigravity-delegate@ai-research-skillsInstall all 5 plugins in one go:
bash scripts/install-all.shOptional Python runtime:
pip install research-hub-pipeline
research-hub setup --persona researcher
research-hub doctorFor non-Claude-Code hosts (Codex CLI, Cursor, Windsurf, Gemini CLI, generic API clients): load the repo's SKILL.md into the host's skills/rules directory or inline it into the prompt. Hermes 0.13.0 is verified end to end for literature-triage-matrix; OpenClaw is structurally compatible but not release-grade verified.
How do you use this agent?
After installing, describe your goal in natural language in Claude Code, e.g.:
Build a literature triage matrix for human-in-the-loop research agents. Verify identifiers, preserve contradictions, and show me the evidence gaps before recommending next steps.Or name the skill explicitly:
Use `literature-triage-matrix` to compare these 5 papers.With the Python runtime, run the automated literature pipeline:
research-hub auto "agent-based modeling" --max-papers 3 --no-nlmCautions: enable Zotero's local API in the desktop app before any write operations, and always export a Zotero backup first (File → Export Library → Zotero RDF)—zotero-library-curator is read-only, but zotero-skills can merge, delete, and rebind items. Chinese-language literature and Chinese output are both supported.
What are this agent's strengths and limitations?
- The 17 skills hand off via YAML schemas rather than free text, with status enums and anti-leakage rules that structurally block evidence-free claims from shipping as supported.
- Portability is audited: the universal SKILL.md layer scored 11/14 zero-edit portable in the 2026-05-10 audit, with a recorded Hermes 0.13.0 end-to-end install verification.
- Five explicit human gates are built into the correctness boundary—destructive actions (external writes, release authorization) pause while read-only reversible work proceeds automatically.
- The maintainer publishes per-skill time/turn expectation ranges and real dogfood artifacts (test-corpus), making cost estimation concrete.
- Built and tested by one graduate-student researcher; not corpus-scale validated, with domain bias toward water resources and agent-based modeling—unvalidated for social sciences, ML, or clinical writing.
- A Claude Code marketplace install guarantees nothing about other hosts' live behavior; OpenClaw is not release-verified, and some hosts need cosmetic <skill-root> path edits.
- Full capability is spread across multiple source repos (research-hub, academic-writing-skills, zotero-skills, etc.); this catalog is only an index, so debugging crosses repos.
- CI performs no claude plugin install round-trip; the actual install and trigger path is verified by the maintainer between releases.
How does this agent compare with similar options?
The README situates the project within the agentic AI learning roadmap (awesome-agentic-ai-zh) but names no direct competitor; unlike prompt-template collections, it enforces inspectable YAML-schema handoffs and human gates, with skill sources split across standalone repos.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AI Research Skills This agent | 77 · Some gaps | Agent plugin / skillFree + model costs | ★ 389 | today | Python | Claude Code |
| LLM Agents Papers | Insufficient evidence | Library / SDKFree | ★ 3.1k | 3d ago | — | — |
| COMPASS Skills | 85 · Good | CLIFree | ★ 758 | today | Python | Codex · Claude Code |
| SLICC Browser-Native Agent | 77 · Good | CLIFree + model costs | ★ 34 | today | TypeScript | Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Catalog repo performs no sensitive operations itself, but declares privilege and confirmation boundaries via mutation_class (external-write/mixed), human_gates, and an explicit Zotero-backup-before-CRUD warning. Deductions: runtime pip dependency (research-hub-pipeline) is not version-pinned or manifested; no sensitive-data guidance (research notes may contain unpublished material; no privacy/redaction handling); rollback relies on manual user Zotero export, with no automated snapshot or recovery path. Publisher is unverified; provenance rests on self-authored links to sibling repos.
Self-consistency is strongly evidenced: pytest enforces catalog/marketplace/README/glossary agreement (17 skills, T1/T2 counts, unified install command), with fail-closed negative schema tests. Deductions: failure messaging is only visible in the monthly health workflow's checker_crash fallback; skill-level error handling is not statically verifiable here; availability depends on external CLIs (codex, agy, Zotero desktop API), only partially version-verified (Hermes 0.13.0).
Per-host install paths (Claude Code/Codex/Cursor/Gemini/OpenClaw/Hermes) are enumerated with host-specific caveats; trigger precision is served by a 17-row trigger map plus an explicit fallback ('name the skill'); each persona gets a starting point. Deductions: capability boundaries rest mainly on the Limitations section; actual SKILL.md bodies live in sibling repos, so depth of boundary declarations cannot be verified; OpenClaw is flagged as not release-verified.
Information architecture is excellent: pipeline docs, skill directory, bilingual READMEs, glossary, per-stage deliverable examples; a Limitations section exists and is test-guarded; MIT license is complete. Deductions: CHANGELOG is referenced but absent from the evidence, so versioning cadence is unverifiable; maintenance responsibility points to a single unverified author across five repos, raising orphan-repo risk; FAQ/example file contents are not provided.
Outputs have concrete shape (YAML schemas with anti-leakage rules, triage matrices, decision dossiers); marginal value targets real pain points (P1–P5: context re-explaining, hallucinated citations); the cost table honestly labels itself 'maintainer-observed ranges, not benchmark numbers'. Deductions: output usability is supported only by descriptions and example links; the executable skill bodies are outside this repo, so actual output quality cannot be statically confirmed.
Fact/inference separation is well done: verification status, tiers, verified_on dates, and verification notes are structured and schema/test-enforced, and timing claims are explicitly non-benchmark. Deductions: cross-source corroboration is weak — all 'verification' is self-reported by the same author (dogfood runs, upstream PRs); the claimed 23-test end-to-end suite is not visible in these files.
- Publisher identity is unverified; all 'verified' statuses in the catalog are self-reported by the author with no independent reproduction.
- zotero-skills can perform destructive writes (merge duplicates, delete items); manually export a Zotero backup first — the skill will not back up for you.
- Runtime Python dependency (research-hub-pipeline) and external CLIs (codex, agy) are not version-pinned; review them before installing.
- The catalog is a registry, not the implementation: the 17 SKILL.md bodies live in other repos, and this repo's tests cannot cover their actual behavior.
- No sensitive-data guidance: unpublished research data processed via Claude/Codex/Antigravity is subject to your institution's confidentiality and licensing rules.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.7.7 Catalog 1.7.7: writing and delegation integrations
Release notes excerpt
Catalog 1.7.7 collects the accepted writing and delegation integrations since v1.7.6. - academic-writing-skills 1.3.7 -> 1.3.17: reader continuity and limitation triage; contextual clarity and independent-reader/rubric guidance; Results, terminology, provenance and coverage safeguards; workflow graphics and supported-client documentation. …View on GitHub ↗ -
v1.7.6 v1.7.6 — Academic writing skill 1.3.7
Release notes excerpt
Academic-writing skill distribution sync - Marketplace metadata 1.7.6 now records academic-writing-skills 1.3.7 and its reviewed source merge. - English and Traditional Chinese version/snapshot references are aligned; other plugin versions, names, source refs, skill counts, and existing unreleased entries are preserved. - The source skill strengthens contextual diagnosis of repetitive procedural sentence frames and empty connectives, while retaining necessary sequence and deliberate parallelism. …View on GitHub ↗ -
v1.7.4 v1.7.4 — Academic writing skill 1.3.6
Release notes excerpt
Publishes the existing academic-writing-skills marketplace entry at 1.3.6 and refreshes its accepted-source snapshot to merge 8a5de8324c9761b89c681745967833f08e2c3061. The upstream release adds first-draft reader progression, author-defined comparison framing, caption/body allocation, saved-file layout and editing identity checks, and optional exact-artifact handoff coverage validation. A clean language scan or complete coverage record is not editorial/scientific certification. …View on GitHub ↗