Seven agent skills for AI-for-Science research — turn a research direction into
literature surveys, runnable experiments, publication-grade papers, and integrity
audits, with every citation, number, and figure traceable to its source.
English · 中文
Contents
The skills
| Skill |
Role |
Primary output |
| ai4s-agent |
Runs the four skills below in order |
the full package |
| research-explorer |
Explore topics from a broad direction |
research_exploration.md, topic_matrix.md, literature_pre_survey.md |
| literature-survey |
Write a literature survey |
6–20 pp PDF, 60+ real citations, LaTeX source, taxonomy figures |
| experiment-suite |
Build an experiment package |
design doc, runnable code, results.json with provenance, figures, report |
| paper-writer |
Write a research paper |
8–14 pp PDF, 200+ citations, 4–8 figures, tables |
| mindmap-render |
Render a mindmap |
image from a topic_matrix.md (Python script) |
| integrity-auditor |
Audit a paper’s integrity |
image / numerical / logical findings, 4-level evidence grading, audit_report.md |
Each skill is a folder with a SKILL.md plus its own references, templates, and
tools. MIT-licensed; works with Claude Code, Cursor, Codex, and Aider.
How they connect
direction
│
▼
[1] research-explorer ──▶ pick one $TOPIC
│
├──▶ [2] literature-survey → survey PDF + bibliography.bib
├──▶ [3] experiment-suite → results.json + figures/
└──▶ [4] paper-writer → paper PDF (reuses [2] and [3])
integrity-auditor ──▶ audits any paper: external PDF / DOI / arXiv, or [4]'s output
ai4s-agent runs steps 1–4 in order. Skills pass work to each other through a
shared slug and the path output/<skill>/<slug>/latest/.
Authenticity
The focus of the project. Every skill enforces:
| Principle |
In practice |
| Real citations |
Every BibTeX entry links to a URL the agent fetched in the same session; none from memory. |
| Labelled numbers |
Every number is marked measured, simulated, or illustrative; simulated values are never reported as measured. |
| Runnable experiments |
experiment-suite outputs runnable code and a results.json with provenance. Supply real results and they replace the simulated ones; the “simulated” disclosure is then removed. |
| Resumable runs |
Long tasks save progress after each step and continue from the last checkpoint, so a reported “done” reflects completed work. |
| Publication layout |
booktabs tables, [!t] floats, ~\cite{}; vector-PDF figures with embedded fonts and defined color palettes. |
| Review disclosure |
Every generated document states that domain-expert review is recommended. |
| Integrity checks |
integrity-auditor inspects a paper for image, numerical, and logical problems and grades the evidence. |
Example
A complete run from experiment-suite + paper-writer: “Learning the Burgers
Solution Operator with a Fourier Neural Operator” — an 8-page paper backed by code
the agent wrote and ran. Full artifact in examples/fno-burgers/
(paper, code, results.json, report).
The 8-page paper (first 6 pages) — click any page for the full PDF.
- Real code, really run —
model.py is a 1-D FNO; the full study runs in ~35 min on a laptop CPU.
- Measured results — FNO 4.41% rel-L2 vs MLP 21.63% and CNN 67.69% (3 seeds); zero-shot super-resolution holds 4.2–4.7% from grid 128 to 1024.
- Real citations — 22 references, each traceable to its source.
Every number is measured (provenance in results.json); the paper states it was
AI-generated and recommends domain-expert review.
Install
Run the installer from the project you want the skills in:
git clone https://github.com/ai4s-research/ai4s-skills
cd /path/to/your-project
/path/to/ai4s-skills/install.sh # all skills → ./.claude/skills
/path/to/ai4s-skills/install.sh paper-writer # or specific ones
SKILLS_DIR=~/.claude/skills /path/to/ai4s-skills/install.sh # global instead
To install by hand, copy any skills/<name>/ into ~/.claude/skills/ (global) or
<project>/.claude/skills/ (project).
Usage
In Claude Code:
Use the literature-survey skill to write a survey on <your topic>.
With Cursor, Codex, or Aider, point the agent at the skill file:
Read skills/literature-survey/SKILL.md and its references/, then produce the survey
for "<your topic>" as specified.
Each SKILL.md directs the agent to read its references/ first; those files hold
the procedures for bibliography expansion, figures, layout, and quality checks.
Repository layout
ai4s-skills/
├── skills/
│ ├── ai4s-agent/ SKILL.md + references/
│ ├── research-explorer/ SKILL.md
│ ├── literature-survey/ SKILL.md + references/ + templates/survey/
│ ├── experiment-suite/ SKILL.md + references/ + figure_examples/
│ ├── paper-writer/ SKILL.md + references/ + templates/paper/
│ ├── mindmap-render/ SKILL.md + scripts/ + tests/
│ └── integrity-auditor/ SKILL.md + references/ + forensics_tools/ + templates/ + tests/
├── tools/validate_skills.py structure / frontmatter validator (run in CI)
├── install.sh
└── .github/workflows/ci.yml
Each SKILL.md carries YAML frontmatter (name, description) so an agent can
find and route to it.
Small, single-purpose scripts the skills call. Each directory has its own
requirements.txt.
skills/integrity-auditor/forensics_tools/ — image duplication / ORB matching, panel splitting, channel checks, magnitude (Benford-style) consistency, decimal matching, spreadsheet aggregate consistency.skills/experiment-suite/figure_examples/ — a matplotlib style kit (style_kit.py) and worked figure examples.skills/mindmap-render/scripts/ — generate_mindmap.py.
Contributing
A new skill needs:
skills/<name>/SKILL.md with name and description frontmatter (name = folder name).- Optional
references/, templates/, and tools.
- No
import anthropic / import openai.
python tools/validate_skills.py passing (CI runs it on every PR).
See CONTRIBUTING.md and the Code of Conduct.
Citation
If you use AI4S Skills in your research, please cite it:
@software{ai4s_skills,
author = {{The AI4S Skills Contributors}},
title = {AI4S Skills: open-source agent skills for AI for Science},
year = {2026},
version = {0.1.0},
doi = {10.5281/zenodo.21297455},
url = {https://github.com/ai4s-research/ai4s-skills},
license = {MIT}
}
The DOI above cites all versions; GitHub’s “Cite this repository” button
(generated from CITATION.cff) provides the same reference
in APA and BibTeX.
License
MIT.
Outputs are drafts. Review by a domain expert is recommended before any citation,
submission, or decision. Verify numbers, citations, and claims.
Acknowledgments
Thanks to linux.do — a vibrant tech community where this project is shared and discussed.
评论