Regression ratchet
Reusable public SKILL.md workflows for AI agents
npx -y skills add selamy-labs/agent-skills --skill regression-ratchetAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when fixing a bug, incident, or regression. In the same change, add the guard that would have caught it — matched to the failure class — so the problem can recur at most once. A fix without its ratchet is half-done.
SKILL.md
2.1 KB, 441 tokens by cl100k_base, as published. Nobody here has run it
Regression Ratchet
A bug that can silently recur is a bug you will fix twice. The ratchet makes the fix monotonic: the same change that fixes the problem also adds the guard that would have caught it. Quality only goes up — the class you just closed cannot reopen unnoticed.
Same change, not "later"
A follow-up "add tests" task rots while the bug is now invisible (it's fixed today). The guard ships in the fix's own PR, or the fix isn't done. If you're tempted to defer it, you're choosing to debug this again.
Match the guard to the failure class
The guard is whatever would have caught this failure — not always a unit test:
| Failure class | Ratchet |
|---|---|
| Logic bug in code | a unit/behavior test reproducing it |
| Wrong artifact / container / file shape | a structure or build test asserting the shape |
| Config / policy / permission gap | a policy or lint check that fails on the bad config |
| Integration / runtime edge | an integration test or a runtime assertion/precondition |
| Genuinely expensive to test | an alert/monitor that fires on recurrence |
Prove it actually ratchets
The guard must fail on the pre-fix state and go green after — run it red→green. A test that stays green with the bug still present guards nothing; it's theater. If you can't make it fail without the fix, you haven't reproduced the failure yet.
One incident, one ratchet
Add the minimal guard for this class — don't bolt on a speculative suite. The discipline is coverage of what broke, not maximalism. Breadth comes from many incidents each leaving one guard behind, over time.
The test
For the bug you just fixed, answer: "If this exact thing happened again, what would catch it automatically — and is that guard in this PR?" If the answer is "nothing" or "a separate task," the ratchet is missing.
Gives 0 of the 12 instructions most data analysis skills give in 441 tokens
Counted across 286 of the 286 authors here whose files we hold, read 2026-08-06
- use excel formulas instead of hardcoded calculated valuesin 35 of 286, across 7 files
- match existing template conventions when modifying filesin 35 of 286, across 7 files
- document sources for all hardcoded valuesin 35 of 286, across 7 files
- write minimal concise python codein 35 of 286, across 7 files
- place all assumptions in separate assumption cellsin 32 of 286, across 5 files
- apply industry-standard color coding to financial modelsin 31 of 286, across 5 files
- format years as text stringsin 30 of 286, across 3 files
- recalculate formulas using recalc.py after modificationsin 30 of 286, across 3 files
- format negative numbers using parenthesesin 30 of 286, across 3 files
- fix all identified formula errors before finishingin 27 of 286, across 1 file
- use colorblind-safe palettesin 19 of 286, across 12 files
- Name tests after the prevented bugin 13 of 286, across 8 files
Said here and by no other author read
- add the guard in the fix's own change
- match the guard to the failure class
- prove the guard fails without the fix
- add the minimal guard for the failure class
- reproduce the failure before writing the guard
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.