prove-it

SKILL.md

Claude generates code by pattern-matching on training data. Something can look syntactically perfect, follow best practices, and still be wrong. The model optimizes for "looks right" not "works right." Verification is a separate cognitive step that must be explicitly triggered. This skill closes the loop between implementation and proof.

1. Generation vs Execution I generate code but don't run it. I predict what it would do based on patterns. My confidence comes from "this looks like working code I've seen" not from "I executed this and observed the result."

2. Training Signal Mismatch My training optimizes for plausible next-token prediction, not outcome verification. Saying "Done!" feels natural. Verifying feels like extra work.

在宣布任何任務完成之前，實際驗證結果。運行代碼。 測試修復。檢查輸出。克勞德的訓練優化了看似合理的外觀 輸出，未經驗證的正確輸出。該技能強制執行驗證步驟 這不是自然而然的。沒有證據就沒有勝利。 來源：elliotjlt/claude-skill-potions。

查看原文