agentleFS
Jinsong-ZhouJinsong-Zhou/bio-research-skills/skills/code-reproduction/SKILL.md

code-reproduction

Skill0 starsChanged 4 months agoMITView on GitHub

This skill should be used when deciding whether a research code repository can actually be run here, before committing time to it: licence layers read separately (code, weights, data, third-party), the credentials, downloads and GPUs it will demand, the install failures its build script swallows, all gated against this machine. Inference by default; training only when asked. Triggers on "can I reproduce this", "can I run this repo", "复现这篇论文的代码", "what do I need to run this", "will this run on my machine", "what licence are these weights", "is this repo usable", "why won't this install", or any paper-plus-repository pair where the question is feasibility rather than science.

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. JunMA98/Computer-science-claude-skills★ 3same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository2 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    This skill does not execute the repository: No environment is built, no weights are downloaded, no model is run.

    line 50
  2. Very specific:

    Do not assume the default branch is main.

    line 68
  3. Distinctive:

    If the user already has a checkout, survey that instead: what is on their disk is what they will run.

    line 74
  4. Distinctive:

    If any exist, read them and prefer them for anything about how to run the pipeline.

    line 83
  5. Distinctive:

    Always give the user three things, in this order:

    line 137
  6. Distinctive:

    The verdict and the one thing that decides it: Not a list — the single gate that matters most.

    line 139

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Code reproduction

A link at the end of a paper is not a reproduction. Between the two sit a licence that may forbid the intended use, weights behind a login, a dataset that shipped as a list of accession numbers, a CUDA build the manifest does not mention, and a GPU this host may not have. Each is cheap to check and expensive to discover — the expensive discovery being the one that happens four hours in.

This skill answers one question: is it worth starting? It does not run the model. When the answer is yes, it hands over to a skill that does.

Scope

Inference is the default target. Reproducing inference — released weights, released data, a published number — is what people attempt, and the obstacles are obstacles of access. Training is a different exercise with different obstacles, and it is opt-in: pass --target training only when the user has asked for training specifically. Never quietly widen the scope; a report that says ok for inference says nothing about training, and gate.py prints that sentence unprompted.

This skill does not execute the repository. No environment is built, no weights are downloaded, no model is run. Everything below reads files and inspects the local machine.

Procedure

1. Establish the target

Inference unless the user said training. If they said "reproduce the paper" without qualifying it, that is inference — say so in one line, so the assumption is theirs to correct.

2. Get the repository

git clone --depth 1 https://github.com/OWNER/NAME /tmp/repro/NAME

Do not assume the default branch is main. If the clone 404s, or if a documented file is missing, check what the default actually is — research repositories sit on dev, master, public and release at least as often. survey.py records the branch it read so the report cannot be mistaken for a different one.

If the user already has a checkout, survey that instead: what is on their disk is what they will run.

3. Look for the repository's own instructions first

ls .claude/skills/*/SKILL.md AGENTS.md CLAUDE.md .cursor/rules 2>/dev/null

If any exist, read them and prefer them for anything about how to run the pipeline. Whoever wrote them knows it better than any general survey. This skill then covers what such files almost never do: the licence layers, the credentials, and whether this host qualifies. survey.py reports them as handoff.repo-ships-guidance and gate.py puts the pointer above the gates.

4. Survey the repository

python3 scripts/survey.py /tmp/repro/NAME --out survey.json --format text

Six groups of checks — licence, weights, data, env, hardware, handoff — each producing findings that name the file and line they came from. Anything a check could not settle is recorded under inconclusive rather than dropped.

5. Probe this machine

python3 scripts/probe.py --from-survey survey.json --disk /path/for/weights --out probe.json --format text

--from-survey makes the probe check exactly the credentials and hosts the repository asked for. Without it those come back unknown, which is honest but less useful. --disk should point where the weights will actually land, not the current directory.

The probe reads whether a credential is set. It never reads the value, so probe.json is safe to keep alongside a reproduction log.

6. Gate

python3 scripts/gate.py --survey survey.json --probe probe.json --target inference --out REPRODUCTION.md

Exit status is 1 when the verdict is blocked or unknown, 0 otherwise.

Reading the verdict

VerdictWhat it meansWhat to do
blockedA requirement is stated and this host does not meet itStop. Report the specific gates and what would clear them
unknownA requirement or a capability could not be determinedResolve it before committing time — an unknown is not a pass
degradedIt will run, with a documented problem aheadProceed, having read the problems
okEvery stated requirement found was metProceed

unknown ranks worse than degraded on purpose. The failure this skill exists to prevent is a report that reads clear because a check quietly found nothing, so a survey that could not find a VRAM figure must not be reported as a model that fits.

Reporting

Always give the user three things, in this order:

  1. The verdict and the one thing that decides it. Not a list — the single gate that matters most. "Blocked: no CUDA device, and the pipeline is single-GPU CUDA throughout."
  2. The licence summary, every time, even when the verdict is ok. Name each layer and its terms: code, weights, data, third-party. This is the only finding that can invalidate work after it succeeds, so it does not get filed under "details". See references/license-layers.md.
  3. What it would take to clear the blocks, concretely. "A single A100 or H100 with 40 GB, Ubuntu 22.04+, an HF_TOKEN, and 200 GB of disk" is actionable; "you need a better machine" is not.

Then, if the user wants to go ahead, hand over — references/handoff.md names the skills that run repositories and their licences. Do not reimplement what they do.

When the verdict is blocked

Say what is blocked and stop. Do not offer to "try anyway" — the point of the check is that trying anyway is the expensive path. Two blocks are worth special handling:

  • No GPU on this host. The repository is not unreproducible; this machine is unsuitable. Say which card would do, and what it would cost to rent one. Offer to survey against a different host later: survey.json is portable, so only probe.py and gate.py need re-running there.
  • A restricted licence. This blocks use of the output, not execution. Be precise about which: an academic user reading "blocked" on a research-only weights licence may be perfectly entitled to proceed.

When the verdict is unknown

Name the specific unknown and how to settle it. Most resolve in one command — nvidia-smi, ldd --version, df -h, or reading one paragraph of a model card. Do not average an unknown into an overall impression.

What this cannot see

State these plainly rather than letting silence imply coverage:

  • Whether the code is correct, or whether the published numbers are reproducible at all. This checks access, not science. Use paper-deep-reading to judge whether the claims were supported in the first place.
  • Undocumented requirements. A model that needs 60 GB while the README says 24 will pass this gate and fail at run time. The survey reads what the repository says about itself.
  • Licence meaning. The layers and their families are identified; what a clause permits in a given jurisdiction is a question for a human, and for anything commercial, a lawyer.
  • Anything behind a login. Whether an account has been granted access to a gated model is not visible from outside it.

References

  • references/license-layers.md — the four layers, how to read each, and the traps that recur
  • references/repro-hazards.md — the catalogue of failure modes each check hunts for, with worked evidence
  • references/handoff.md — which existing skill to hand execution to, and under what licence

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.