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
- JunMA98/Computer-science-claude-skillstheir own version, first seen Mar 11, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 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.
- literature-trackingSkill
- paper-deep-readingSkill
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
- Distinctive: line 50
This skill does not execute the repository: No environment is built, no weights are downloaded, no model is run.
- Very specific: line 68
Do not assume the default branch is main.
- Distinctive: line 74
If the user already has a checkout, survey that instead: what is on their disk is what they will run.
- Distinctive: line 83
If any exist, read them and prefer them for anything about how to run the pipeline.
- Distinctive: line 137
Always give the user three things, in this order:
- Distinctive: line 139
The verdict and the one thing that decides it: Not a list — the single gate that matters most.
≈ “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.
Files
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/NAMEDo 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/nullIf 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 textSix 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.mdExit status is 1 when the verdict is blocked or unknown, 0 otherwise.
Reading the verdict
| Verdict | What it means | What to do |
blocked | A requirement is stated and this host does not meet it | Stop. Report the specific gates and what would clear them |
unknown | A requirement or a capability could not be determined | Resolve it before committing time — an unknown is not a pass |
degraded | It will run, with a documented problem ahead | Proceed, having read the problems |
ok | Every stated requirement found was met | Proceed |
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:
- 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."
- 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. - 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.jsonis portable, so onlyprobe.pyandgate.pyneed 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-readingto 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
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
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.

