principle-reproduce-from-clean-state
Skill1 starsChanged 15 days agoMITView on GitHub
Apply before reporting a result, handing off a notebook, or merging an analysis. A result counts only when it re-runs from a clean kernel or process, from the source data, with pinned dependencies (uv.lock) and fixed seeds. Hidden notebook state and cell order are bugs.
In this repository50 other files it gives its agents.
- analyst-modeSkill
- architectSkill
- arenaSkill
- astack-helpSkill
- benchmark-checklistSkill
- blast-radiusSkill
- catch-upSkill
- correctSkill
- create-formal-verification-skillSkill
- figure-it-outSkill
- howSkill
- interrogateSkill
- in-your-own-wordsSkill
- maintain-formal-verification-skillSkill
- no-commentsSkill
- plain-wordsSkill
- principle-attack-the-premiseSkill
- principle-boundary-disciplineSkill
- principle-build-the-leverSkill
- principle-define-the-questionSkill
- principle-encode-lessons-in-structureSkill
- principle-experience-firstSkill
- principle-explain-the-numberSkill
- principle-fix-root-causesSkill
- principle-foundational-thinkingSkill
- principle-guard-the-context-windowSkill
- principle-laziness-protocolSkill
- principle-make-operations-idempotentSkill
- principle-minimize-reader-loadSkill
- principle-model-the-domainSkill
- principle-name-the-estimandSkill
- principle-never-block-on-the-humanSkill
- principle-prove-it-worksSkill
- principle-respect-the-grainSkill
- principle-sequence-verifiable-unitsSkill
- principle-simulate-the-methodSkill
- principle-subtract-before-you-addSkill
- principle-test-behavior-not-implementationSkill
- recallSkill
- reflectSkill
- retail-analyticsSkill
- sharpenSkill
- show-me-your-workSkill
- statechartSkill
- swarmSkill
- tddSkill
- teachSkill
- technical-writingSkill
- unslopSkill
- whySkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
Read from the source data path, not from an intermediate you made by hand.
- Distinctive: line 18
Keep raw data, credentials, and notebook outputs out of commits.
- Distinctive: line 19
Run it a second time and compare.
≈ “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
Reproduce from Clean State
A result counts only when it can be reproduced from scratch. Restart the kernel or open a new process, run from the top, and read what comes out. On real data the human does the run, from a run request, and pastes the summary back. A number that exists only in a warm kernel is not a result.
Notebooks keep state across cells. A variable from a deleted cell, a file overwritten earlier, or cells run out of order can produce numbers that no one can rebuild. Saved outputs show what was true once, not what the code does now.
- Re-execute top to bottom in a fresh process. Use "Restart and Run All" in the notebook tool, or run the converted script with
uv run. - Pin the environment with
uv.lock, commit it, and run throughuv runso the versions match. Do not rely on whatever happens to be installed. - Set and record a seed for any sampling, bootstrap, or model fit. Pass the seed as a parameter, not as a value hidden in a cell.
- Read from the source data path, not from an intermediate you made by hand. If a cache is used, its key includes the source version, such as a date or a hash.
- Pin the input. Every run takes an explicit as-of timestamp, partition, or snapshot id, and filters the source to rows that existed at that point, for example
ingested_at <= as_of. Write the as-of value into the output next to the command that produced it. Without a pinned input, two clean runs days apart can differ for reasons that are not bugs. - Cell order is a bug. If a top-to-bottom run gives a different result from the saved one, the notebook is wrong. Fix the order, or move the logic into functions.
- Keep raw data, credentials, and notebook outputs out of commits. Commit the code and the lockfile. The number lives in the report, next to the command that produced it.
- Run it a second time and compare. Two clean runs that match are the minimum evidence. Ask for both runs in one request, and have the script print a checksum of its outputs so the two pastes compare at a glance.
You skipped this when a result exists only in a warm kernel, when a saved notebook output is quoted as the answer, or when a rerun with the same inputs gives a different number.
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.

