wtf is skill-extractor?

skill-extractor scans through the transcripts of your coding agent sessions, tools like Claude Code or OpenAI Codex CLI, and pulls out reusable patterns it noticed you repeating: things like always running tests before saying a task is done, or avoiding a mistake that caused a previous session to fail. Instead of a general purpose memory feature that quietly changes an agent's behavior, this tool writes each discovered pattern out as a plain markdown skill file that you can read, edit, or delete yourself. The author built it after noticing they kept repeating the same corrections to their coding agent session after session, since those lessons were forgotten once each conversation ended. Every candidate skill it finds is scored by how useful and trustworthy it seems, weighted by whether the session it came from actually succeeded or failed, and only strong, new candidates are shown to you for review through a VS Code panel, a Claude Code slash command, or a plain command line tool. The project is explicit about its security tradeoffs. It runs on your own machine and reads private transcripts that may contain code, infrastructure details, and occasionally secrets, redacting common secret patterns like API keys and tokens before sending anything to a language model. By default it uses your existing Claude subscription to do the mining, but it can be pointed at a fully local model instead for zero external data transfer. Every skill it wants to install also runs through a safety check looking for risky instructions, and you must explicitly approve any flagged skill before it takes effect. It supports multiple agent sources sharing one review queue, six categories of skills such as workflows and guardrails, and is designed to make no model calls at all when there is nothing new to mine, keeping ongoing cost low. The full README is longer than what was shown.