verify-before-ship
Skill10 starsChanged 7 days agoMITView on GitHub
Principles for handing over, verifying, and approving any change (code, documents, instructions) before it ships. Use to do or review such work, or on /verify-before-ship.
In this repository35 other files it gives its agents.
- AGENTS.mdAGENTS.md
- design-principlesSkill
- experimentationSkill
- groupingSkill
- idea-qualitySkill
- info-foragingSkill
- investingSkill
- listeningSkill
- misfitsSkill
- motivationSkill
- needs-to-signalsSkill
- optionalitySkill
- perspectiveSkill
- purposeSkill
- testingSkill
- resume-deep-diveSkill
- review-resumeSkill
- inspiring-ideaSkill
- next-goalSkill
- side-quest-recommendSkill
- still-want-itSkill
- find-the-argumentSkill
- make-presentation-slideSkill
- agent-evalsSkill
- behavior-viewSkill
- black-boxSkill
- clean-arch-domainSkill
- mutation-testingSkill
- structure-viewSkill
- test-oracleSkill
- unit-testsSkill
- user-storiesSkill
- white-boxSkill
- negotiationSkill
- readableSkill
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 27
Works: every required check passed on the exact version the owner gets.
- Distinctive: line 29
If any of the three is missing, it is not done.
- Distinctive: line 56
Do not shape the work to fit its checker: When a measure becomes a target, it ceases to be a good measure.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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
/verify-before-ship
Ship only work the owner can accept after one read, with evidence and approval for the exact version that ships. Use this skill to do the work or to review it; in a review, name each breach once, with its fix, most important first. Any report you write follows principle 1 too.
Principles
- Done means the owner can accept it after one read. Reading the work
once, the owner finds three things right:
- Core idea up front: the most important, highest-level idea comes first. Principles come before plans. Definitions and support come later.
- Reads easily: plain words, no needless words, and no term the owner must look up.
- Works: every required check passed on the exact version the owner gets. If any of the three is missing, it is not done. Every check that can run is run before handover; skipping one and saying so instead is not allowed. Before calling a check impossible, look for another way to run it, and say what you tried. A check that has not run leaves the work not done, whichever path follows: it stays not done until the check runs and passes on this version, or until the owner, told plainly which check could not run and why, decides. A "not done" verdict that calls for an edit also calls for every required check to run again on the edited work, not only the one that failed (principle 3).
- The builder and the verifier are different parties. Anyone who helped build the work cannot be its verifier. The verifier sees only the task, the work, and the criteria, and works out the problem from them, not from the builder's account. Keep the builder's reasoning and expected answers out of view (closed-book) until the verdict is in; after that they may help explain differences. The verifier reports findings to the owner without the builder's approval.
- Evidence and approval hold only for the version they were given on. Editing the work, or anything it depends on, cancels both, whatever its size or reason. That includes a fix made to satisfy another principle here, and any fix you prescribe: run every required check again on the changed version, then seek approval again. A change only to how the work is described resets no check, but the description must still match the evidence: remove or check any claim the evidence does not support. Time passing changes nothing by itself; when the work depends on something that changes with time, that change counts as an edit to what the work relies on. Never report a check as passed unless it ran on this version. When handing over after a change, show what changed and which checks ran again.
- Do not shape the work to fit its checker. When a measure becomes a target, it ceases to be a good measure. Do not copy the checker's wording or expected answers into the work; fix the behavior they point at. Keep a held-out check that never guides changes and runs once, at the end, on the version that ships; lead with its result as the one to expect. Seeing a held-out result uses it up: if the work changes after that, report its later results as tuning scores and replace it with a fresh one. Call any score on checks that shaped the work a tuning score, say it overstates real performance on unseen cases, and give the gap to the held-out score. A large gap means the work fits the checker, not the task. For a measure tracked over time, also track one or more measures never treated as a target, and watch how far the two drift apart; swapping in a new target only creates the next target.
- Make the smallest change that solves the whole task. Each part must earn its place. Leave out work the task does not need, including for needs that do not exist yet; you may offer it to the owner as a later option. A change that leaves part of the task unsolved is not smaller, it is unfinished.
- Nothing ships until the owner explicitly approves it. Green checks and other reviewers' sign-offs support the owner's decision; they do not make it. Work stays a draft until the owner, or someone the owner named, approves this exact version after its checks have run.
Terms
- Owner: the person who decides whether the work ships.
- Check: any test, review, or audit whose result is evidence about the work. A required check is one the task, the criteria, or the owner calls for.
- Version: one exact, identifiable state of the work and of everything it depends on: its inputs and the environment it runs or is used in.
- Description: what is said about the work, apart from the work itself.
- Held-out check: a check kept unseen while the work is shaped, run last.
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.

