Claude Evolves From Its Mistakes

After the same type of mistake repeated 4 times while using Claude Code, a vibe-tuning skill was built to have Claude self-check its own mistakes.

The core idea is that Claude automatically runs a 6-step postmortem whenever it makes a mistake.

• Acknowledge what went wrong
• Retrace its own reasoning to identify the cause
• Find the root cause, not just the symptom
• Propose a type of solution
• Save the fix after user approval
• Generate an enforcement script that actually works

The methodology of this project was itself built inside Claude Code, accumulating real mistakes Claude made during sessions as examples. It emphasizes that Claude wrote everything—SKILL.md, the classification system, and the enforcement scripts.

The most important finding is that memory rules alone aren't enough. Rules can be read but still ignored, leading to the conclusion that they ultimately need to be enforced with pre-execution scripts such as PreToolUse hooks.

The cases that came up during the build are concrete.

• Pushed a personal wiki to a public GitHub repository
• Overwrote the configuration of a running daemon
• Wrote a README intended for a global audience in Russian

All three problems were traced back to a single root cause—prioritizing speed over accuracy in irreversible actions—and were resolved together with a single rule. The repository has been released under the MIT license, and it's introduced as installable with a single line: npx skills add AyanbekDos/vibe-tuning.