multica-ai/andrej-karpathy-skills
This CLAUDE.md turns common agent failure modes into four operating rules: surface uncertainty, minimize code, avoid unrelated edits, and define verifiable completion criteria.
The repository packages **four principles** into one CLAUDE.md: expose assumptions and tradeoffs, write the minimum code, limit edits to the request, and drive work with verifiable goals. A Cursor rule carries the same guidance.
Merge the rules with project-specific instructions, then add concrete tests and acceptance criteria. The strongest operational pattern is **tests-first verification**, giving an agent a target it can check rather than an underspecified command.
The repository packages **four principles** into one CLAUDE.md: expose assumptions and tradeoffs, write the minimum code, limit edits to the request, and drive work with verifiable goals. A Cursor rule carries the same guidance. Merge the rules with project-specific instructions, then add concrete tests and acceptance criteria. The strongest operational pattern is **tests-first verification**, giving an agent a target it can check rather than an underspecified command. These rules intentionally favor caution and may add friction to trivial fixes. They are behavioral instructions, not evidence that every model or task will improve, so judge them by diff scope, clarification timing, and test results.
This turns minimal, scoped, assumption-aware coding into a small reusable instruction layer, with tests and acceptance criteria supplying the verifiable target. It reinforces evidence-led harness design, but also narrows the claim: cautious rules are defaults to evaluate by diff size, clarification behavior, and test outcomes, not proof that every model or trivial task benefits.