Guard important business arithmetic with an independent, dependency-free reference implementation checked against the production code's real output. Use for prices, billing, quotas, scores, forecasts - any figure where a wrong answer still looks like an answer.
Installs into .claude/skills of the current project.
Are you the author of Twin Check?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/fadi88-twin-check)
---
name: twin-check
description: "Guard important business arithmetic with an independent, dependency-free reference implementation checked against the production code's real output. Use for prices, billing, quotas, scores, forecasts - any figure where a wrong answer still looks like an answer."
---
<!-- Generated by tools/sync_adapters.py from skills/twin-check/SKILL.md. Edit the source, not this file. -->
# Twin check
Some bugs produce a number that is wrong and looks right. Tests written by the same mind that
wrote the code tend to share its mistakes. A twin is a second implementation, written plainly,
in a different form, and compared against the real code on real inputs.
## How
- Put the twin under `tools/twins/`, in plain Python with no libraries that do the hard part for
it (a twin that calls the same library proves the library works, not your code).
- Read constants from the production source rather than retyping them, so a changed constant
cannot hide.
- Have the production test suite dump its outputs for a fixed set of inputs, and a check under
`tools/checks/` compare twin and dump field by field, with a stated tolerance.
- A missing dump is a loud SKIP, never a pass.
- The twin's header lists the cases where a wrong answer would still look plausible, and the
check covers each.
Use it where it pays. Most code does not need a twin; money and anything the founder makes
decisions from usually does.
