Auto-Research-Skills Research Toolkit
Browse a curated collection of research skills, systems, and benchmarks spanning the path from research ideas to papers.
- Source repo
- brycewang-stanford/Auto-Research-Skills
- Stars
- ★ 221
- Last updated
- 3d ago
- License
- CC0-1.0
- Primary language
- Python
- FA score
- 65/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A graduate student looking for agent skills for literature review, experiments, or writing can start with the stage-based selections in
catalog/STARTER-KITS.md. - Not a fit if
- Users who only need one lightweight skill, not a broad research toolkit
- Teams that need unique skill names for conflict-free bulk imports
- Source review
- 65/100 · Some gaps
What does this agent do, and when should you use it?
Auto-Research-Skills is a curated hub organized into skills/, systems/, benchmarks/, and lists/, covering reusable skills, end-to-end systems, evaluation benchmarks, and resource lists. The README reports 6,089 skills across 110 repositories, collected as shallow git submodules. The resources target coding agents such as Claude Code, Codex, OpenClaw, and other LLM agents, with examples spanning literature review, experiments, data analysis, paper writing, and peer review. Clone the repository and run setup.sh to initialize its top-level submodules and attempt to fetch nested submodules declared upstream. This is a collection rather than one research agent or hosted service; capabilities, runtime needs, and licensing vary by included project.
The repository groups resources into four directories: skills/ for reusable skills and plugins, systems/ for end-to-end research systems, benchmarks/ for evaluation benchmarks, and lists/ for curated lists. Running ./setup.sh initializes top-level submodules and attempts to initialize nested submodules declared by upstream repositories. The README points users to catalog/STARTER-KITS.md to choose skills by research stage and gives a tools/build_starter_kits.py command to install a stage-specific set. Included projects cover tasks such as literature search and synthesis, experiment code, data analysis, paper writing, and review. The hub itself does not run a unified research workflow or produce a standard research output.
- A graduate student looking for agent skills for literature review, experiments, or writing can start with the stage-based selections in
catalog/STARTER-KITS.md. - A research group comparing autonomous research systems, deep-research tools, and research skills can browse the corresponding directories.
- A researcher using Codex or Claude Code can select a submodule for tasks such as literature review, paper writing, or research coding.
- An empirical social-science researcher looking for resources on identification strategies, data preparation, or regression tables can inspect the highlighted Auto-Empirical-Research-Skills and Stata resources.
- A developer evaluating research agents who needs ML engineering or scientific-programming benchmarks can browse
benchmarks/.
How do you install or deploy this agent?
Git is required, and fetching submodules requires network access. Clone the repository and run setup.sh:
git clone https://github.com/brycewang-stanford/Auto-Research-Skills.git
cd Auto-Research-Skills
./setup.shTo skip nested submodules and fetch only the top-level collection:
ARS_SKIP_NESTED_SUBMODULES=1 ./setup.shThe README does not specify shared runtime dependencies or model credentials; requirements depend on the selected submodule.
How do you use this agent?
Browse catalog/STARTER-KITS.md or the four resource directories, then choose the submodules you need. The README provides this example for installing literature-stage skills:
python3 tools/build_starter_kits.py --install literature --dest ~/.claude/skillsBefore combining skill collections, inspect catalog/collisions.json: different collections may include skills with the same name but different contents. Consult each submodule's own instructions for its subsequent setup and invocation.
What are this agent's strengths and limitations?
- Separate directories for skills, systems, benchmarks, and lists make it easier to browse by resource type or research task.
- Shallow submodules aggregate resources from 110 repositories, and
setup.shcan fetch the collection in bulk. - Stage-based starter kits and a documented skill-collision inventory help users select and combine resources.
- The collection is large, so finding and assembling a suitable set takes work; the README advises against importing every skill at once.
- Included projects come from different upstream repositories, so their runtimes, configuration, and maintenance are not uniform.
- The hub does not provide one research runner, and users must check each submodule's capabilities and license.
How does this agent compare with similar options?
Unlike a repository centered on one research system, this hub collects skills, systems, benchmarks, and curated lists. For a specific capability, use the relevant submodule rather than treating the whole collection as a single runnable agent.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Auto-Research-Skills Research Toolkit This agent | 65 · Some gaps | CLIFree | ★ 221 | 3d ago | Python | — |
| Wisp Science | 71 · Some gaps | Desktop appFree + model costs | ★ 1.1k | today | Rust | Codex · Claude Code · OpenAI API · Claude API |
| AERS: Auto-Empirical Research Skills | 51 · Major gaps | Agent plugin / skillFree + model costs | ★ 4.7k | 3d ago | Stata | Claude Code |
| Open Science Workbench | 52 · Major gaps | Desktop appFree + model costs | ★ 5.4k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes third-party git submodules and recommends setup.sh; SECURITY.md explicitly treats skills as executable supply-chain inputs, says the hub does not sandbox or execute them, and calls scan flags advisory. This supports source and risk-boundary transparency and limited permissions for ordinary cloning. The materials do not show per-action user approval, sensitive-data safeguards, or approval controls for external effects, so those areas receive only partial credit.
The README, SECURITY.md, workflows, and tests broadly align on the catalog, generated reports, and checks; workflows include metadata checks, unit tests, and report validation, while SECURITY.md describes nested-submodule initialization issues. Availability depends on upstream repositories and network cloning, with no availability guarantee. Failure reporting has reproduction and reporting paths but limited user-facing troubleshooting detail, so full marks are not justified.
The README targets researchers and users of coding agents such as Claude Code, Codex, and OpenClaw; it organizes skills, systems, benchmarks, and lists, and includes starter kits and a warning about name collisions. SECURITY.md clearly limits the hub to a read-only catalog layer without sandboxing. Trigger precision is determined by upstream content and is not shown to be uniform across the collection; cross-environment compatibility is only partly documented.
The repository layout, directory index, install commands, and contribution checks are clearly described. The README warns about skill-name collisions, SECURITY.md explains limits and reporting, and LICENSE specifies CC0 1.0. A weekly ranking script and CI provide some maintenance signals, but the materials do not establish a complete release-versioning or maintenance-responsibility policy, so those criteria are not full marks.
The aggregated catalog, search and collision browser, starter kits, and safety and quality reports can help users discover and filter a large set of research skills. The repository is a collection of third-party projects rather than one uniformly runnable research agent. The provided materials do not demonstrate output quality, compatibility costs, or realized screening benefits, so scores remain moderate.
The README's categories and paths, SECURITY.md's responsibility statements, CI workflows, and unit tests provide mutually supporting internal evidence. Tests show selected catalog-tool cases and workflows list validation commands. Many upstream capability and count statements remain README summaries without item-by-item independent verification; this review uses static files only and does not treat promotional descriptions as execution evidence.
- The repository aggregates third-party skills that agents may execute. It explicitly provides no sandbox, and scan results are advisory rather than enforcement. Review selected upstream skills and restrict permissions in the consuming environment.
- The README warns that different collections can contain same-named skills with different contents; check the collision catalog before combining installations.
FAQ
Does cloning the hub make every research feature ready to use?
./setup.sh fetches submodules, but each project's own installation, configuration, and runtime requirements still apply.