ai-for-research 1.0.0

Install the ai-for-research Agent Skills into a coding agent's skills directory.

AI-for-Research Skills

A collection of Agent Skills that cover the research workflow end to end — from first idea to polished paper. Each skill is a self-contained SKILL.md (YAML frontmatter with name + description, then Markdown instructions) that teaches an agent how to do one research task well.

The skills use the open Agent Skills format, so they work with any agent that supports it — Claude Code, OpenAI Codex, Gemini CLI, and others — and the instructions are plain Markdown, so you can also drop a SKILL.md straight into any LLM as context. Nothing in them is Claude-specific.

Everything is tuned for three venue families:

• Top AI/ML conferences — NeurIPS, ICML, ICLR, CVPR, ACL, …
• Computer architecture — ISCA, MICRO, ASPLOS, HPCA — plus the related systems, security, and HPC venues architecture work appears in (MLSys, SC, PACT, ISPASS, IISWC, SOSP, OSDI, USENIX ATC, NSDI, IEEE S&P, NDSS).
• Nature-family science — Nature, Science, Nature [subject], …

The research workflow

The skills group into six stages that flow into one another. The pipeline is iterative, not one-shot — expect to loop back to read more, re-evaluate an idea, or re-run experiments before you write anything up.

   1 · GROUND        literature-survey  ·  paper-read  ·  chalk-talk
        │
        ▼
   2 · IDEATE        idea-brainstorm  ·  idea-evaluate
        │
        ▼
   3 · EXPERIMENT    code-implement  ·  experiment-scaffold  ·  code-clean
        │
        ▼
   4 · ANALYZE       result-generate            (raw results → figures & tables)
        │
        ▼
   5 · WRITE         paper-draft  ·  paper-polish
        │
        ▼
   6 · RELEASE       artifact-create            (cleaned code → reproducible Docker artifact)

Within a stage the two skills pair up: survey the field then read a paper; generate ideas then judge them; build your own harness while running others' code as baselines; draft then polish. Between experiments and writing, result-generate turns raw outputs into the paper's figures and tables, and code-clean tidies the experiment code (yours or others') without changing what it does. Finally, artifact-create packages the cleaned code into a reproducible Docker artifact for release and artifact evaluation. Reading (stage 1) also feeds every later stage — revisit it whenever a gap shows up. chalk-talk sits in stage 1 next to reading but is cross-cutting: turn any concept — one you just read or one you're proposing — into a teaching diagram at any stage.

Entry points — start wherever you actually are:

• New to a topic / need background → literature-survey, paper-read
• Need to teach or visually explain a single concept → chalk-talk (cross-cutting)
• Have a topic but no idea yet → idea-brainstorm
• Have an idea, unsure if it's worth it → idea-evaluate
• Ready to test an idea → experiment-scaffold (with code-implement for baselines)
• Code got messy / cleaning up before release → code-clean
• Have raw results to turn into figures/tables → result-generate
• Have exhibits and need to write them up → paper-draft, then paper-polish
• Packaging code for artifact evaluation / a reproducible release → artifact-create

The skills

Skill	Goal	Reach for it when…
literature-survey	Run a multi-source, multi-phase survey (search → read → synthesize) and produce a cited Markdown report with inline hyperlink citations.	You need the state of the art on a topic, or a related-work section / bibliography.
paper-read	Walk through a paper hierarchically with Keshav's three-pass method (bird's-eye → grasp → deep), stopping when it's not worth going deeper.	You need to read, understand, triage, or review a paper (for a junior researcher).
chalk-talk	Generate a chalkboard-style teaching diagram for one concept — bubbles + arrows, handwritten chalk text — as a layout plan, a paste-ready image-generation prompt, narration, and alt text. (Cross-cutting — usable at any stage, not strictly sequential.)	You want to teach or visually explain a single concept to a class, labmate, or talk audience.
idea-brainstorm	Generate a wide set of research ideas for a topic, then tier them by novelty (T1 paradigm-shift → T5 incremental), verifying novelty against prior work.	You want new directions, or to rank ideas by how genuinely novel they are.
idea-evaluate	Judge an idea through the IEEE Micro Top Picks lens — long-term impact above all — and return a narrative verdict (Top Pick / Honorable Mention / Below the bar).	You need to decide whether an idea is worth pursuing, submitting, or dropping.
code-implement	Reproduce and run someone else's code (GitHub repo or paper artifact) from install to the exact command that generates the result of interest.	You need to get external code working or reproduce a paper's results.
experiment-scaffold	Plan first, get sign-off, then scaffold a runnable experiment framework (a validated "walking skeleton") to test your own idea.	You're standing up experiments from scratch to validate a hypothesis.
code-clean	Clean up existing code — dead code, debug cruft, style, hardcoded values — without changing behavior, behind a safety net (tests or a golden snapshot) with a tool-evidenced deletion log.	You need to tidy/refactor code or strip cruft without breaking it.
result-generate	Turn raw experiment outputs into publication-quality figures and tables — with the analysis (aggregation, error bars, significance, normalization) and honesty (show variance, no cherry-picking) behind them, plus a script that regenerates each exhibit.	You have results and need the paper's figures/tables.
paper-draft	Draft new paper prose for any section, venue-aware — brainstorm the framing first, lead with the contribution, avoid AI-tells.	You have results and need to write them up.
paper-polish	Critique and polish existing paper prose like an area chair — diagnose first, then optional fixes — venue-aware.	You have a draft and want it reviewed, tightened, or de-AI-ed.
artifact-create	Package your code into a reproducible Docker artifact for artifact evaluation — runs code-clean first, bakes in a pinned environment, and verifies it reproduces the results in a clean room, with an Artifact Appendix.	You're preparing an AE submission or a reproducible release.

Pairs that are easy to confuse: experiment-scaffold builds your own harness; code-implement runs others' code. paper-draft generates new prose; paper-polish improves existing prose. artifact-create packages your code into a runnable artifact; code-implement runs someone else's.

Looking for paper review? This repo stops at paper-polish (prose-level, section-by-section critique). For full reviewer-style review of a paper before submission — surfacing what reviewers will flag so you can fix it first to improve quality — see the companion repo UnaryLab/ai-paper-review.

How to use these skills

1. Make them available to your agent

Install the ai-for-research CLI — from PyPI, or editable from a clone — then run it for your agent. It symlinks every skill into that agent's skills directory and is safe to re-run:

pip install ai-for-research                       # from PyPI

# — or editable from a clone (tracks updates via `git pull`) —
git clone https://github.com/UnaryLab/ai-for-research.git
cd ai-for-research && pip install -e .

ai-for-research claude                 # → ~/.claude/skills
ai-for-research codex                  # → ~/.agents/skills
ai-for-research gemini                 # → ~/.gemini/skills

ai-for-research <agent> [project-dir] — pass a project directory to install the skills into that project's agent subdir (e.g. ai-for-research claude ~/my-project → ~/my-project/.claude/skills); with no argument the skills go to the per-agent global default shown above. ($SKILLS_DIR sets an exact directory; CLONE_DIR and REPO_URL are honored too.) Installed editable (-e) from the clone, the command links that checkout so a git pull updates the symlinks; installed from elsewhere it clones the repo on first run.

Codex and Gemini also recognize the shared ~/.agents/skills alias.

Prefer not to pip install? Copy the skills in manually. After cloning, each skill is just a folder under skills/ — copy (or symlink) them into your agent's skills directory:

git clone https://github.com/UnaryLab/ai-for-research.git
cd ai-for-research

mkdir -p ~/.claude/skills            && cp -r skills/* ~/.claude/skills/             # Claude Code
mkdir -p ~/.agents/skills            && cp -r skills/* ~/.agents/skills/             # OpenAI Codex
mkdir -p ~/.gemini/skills            && cp -r skills/* ~/.gemini/skills/             # Gemini CLI

Use cp -r to copy, or symlink (ln -s "$PWD/skills/<name>" <dir>/<name>) so a git pull keeps them updated. Installation path per agent:

Agent	Skills directory
Claude Code	~/.claude/skills/ (global) — or <project>/.claude/skills/
OpenAI Codex	~/.agents/skills/ (global) — or <project>/.agents/skills/
Gemini CLI	~/.gemini/skills/ (global) — or <project>/.gemini/skills/

No runtime? Each skill is just a folder with a SKILL.md — paste the relevant one into any LLM as context (or prepend it to your prompt). It's plain Markdown instructions.

2. Invoke a skill

• Automatically — just ask in plain language. The agent matches your request against each skill's description and When to use section and runs the best fit (e.g. "help me read this paper" → paper-read).
• Explicitly — name the skill: in Claude Code as a slash command (/literature-survey, /paper-polish, …); in other tools however that runtime invokes skills — or simply say "use the literature-survey skill."

3. What to expect

• Some skills are gated on your input: experiment-scaffold won't write code until you approve its plan; paper-read goes one pass at a time and asks before going deeper; literature-survey writes each phase to disk before the next.
• Skills hand off to each other and will suggest the next one where it makes sense.
• A few have helper scripts (e.g. literature-survey/scripts/ for paper search and the citation database); the rest are pure instructions.

Conventions shared across all skills

• Persona & ground rules. Every skill opens with a tailored expert persona plus the same three rules: don't be stupid, don't mansplain, don't ask questions that can be googled (look it up; only ask for what's in the user's head).
• Venue-aware. Guidance, checklists, and examples adapt to the three venue families above.
• No fabrication. Citations link to real papers; reproduced numbers are reported as-is; results trace to the command that produced them.

Repository layout

ai-for-research/
├── README.md                # this file
├── LICENSE                  # MIT
├── pyproject.toml           # Python packaging — `pip install -e .` adds the ai-for-research CLI
├── ai_for_research/         # the ai-for-research CLI package (cli.py)
├── .github/workflows/       # CI — publish.yml (build + publish to PyPI on release)
└── skills/                  # the Agent Skills — one folder per skill, each with a SKILL.md
    ├── literature-survey/   # also has references/ and scripts/
    ├── paper-read/
    ├── chalk-talk/          # cross-cutting: concept teaching diagrams
    ├── idea-brainstorm/
    ├── idea-evaluate/
    ├── code-implement/
    ├── experiment-scaffold/
    ├── code-clean/
    ├── result-generate/
    ├── paper-draft/         # also has references/
    ├── paper-polish/        # also has references/
    └── artifact-create/

Each skill folder contains a SKILL.md; some also carry references/ (deeper guidance read on demand) and/or scripts/.

Acknowledgments

Several skills adapt procedure-as-prompt templates from Todd Austin's toddmaustin/research-prompts (Apache-2.0):

• chalk-talk — adapted from his chalk-talk prompt.
• paper-read — the optional Heilmeier Catechism lens (from heilmeier-extractor).
• idea-brainstorm — the fail-fast "cheapest decisive test" layer (from fail-fast).
• paper-polish — the deadline-triage, fact-faithfulness, page-trim, and presentation modes (from brutal-review, hallucination-detector, orphan-finder, and tough-crowd).
• paper-draft — the VOICE CARD voice-capture/match stage (from writing-voice).

License

MIT © 2026 UnaryLab.