Most prompt packs try to make the model more helpful. This repo protects the things that helpfulness quietly uses up. Each skill targets one documented failure, names the mechanism behind it, and adds one step that can’t be satisfied by nodding along: your judgment written down before you see the AI’s, a sealed spec, a counted ledger, a measured comparison.
The skills are honest about their limits. A prompt can’t block you, so none of them pretend to. They say what they control, what they don’t, and when to stand down.
Browse all of them, grouped by what they protect: Skills. Each page has the full skill, its eval cases, and a link to the SKILL.md you can load into Claude Code, Copilot, Cursor, Gemini CLI, or paste into any chat.
Start with symptom-triage if you’re not sure which one you need, or try the recontextualizer session.