SANITY is a Claude Code “skill” — a drop-in instruction module — built to stop two of the most expensive failure modes of AI-assisted coding: confident hallucinations (made-up functions, APIs, file contents) and insecure code (SQL injection, hardcoded secrets, missing auth). Its motto is blunt: “a wrong answer delivered confidently causes more damage than ‘I don’t know.’”
It works on two axes at once — epistemic integrity (verify before you claim) and security integrity (OWASP Top 10 enforcement). There is a “FORTRESS” framework of mandatory checks (read-before-claim, threat modeling, security audits), a confidence matrix that rates each claim CONFIRMED / INFERRED / UNCERTAIN / BLOCKED, an anti-pattern blacklist, and a single SANITY score — security weighted 60%, epistemics 40%, because a breach hurts more than wasted time. The blocks are hard: OWASP violations and hardcoded secrets cannot be worked around.
Install it with claude skill install <repo>, copy sanity.md into ~/.claude/skills/, or auto-load it from your project’s CLAUDE.md. If you run Claude Code (or similar agents) against production code, it is worth a look — and the README is bilingual (EN/ID). MIT licensed.
source: github.com/verycoolapps/sanity-anti-hallucination-skill
my other side: ivo.blogspot.com
Leave a Reply