nventimiglia/learnskill

LearnSkill is a Python tool that helps AI coding agents, such as Claude, Gemini, Cursor, and Antigravity, get better over time by analyzing their own past work sessions. Rather than calling an external AI service, the agent you are already running performs the analysis itself, so no API key is required. When you invoke the skill by asking your agent something like "analyze my sessions" or "what friction did you see," it reads the conversation and session logs stored on your machine from whichever agent tools you use. Claude Code stores sessions as JSONL files, Gemini CLI saves transcripts in a similar format, Cursor keeps session data in a SQLite database inside workspace storage, and GitHub Copilot logs are stored as JSON files. The skill processes these logs and looks for three types of signal: friction (specific moments where the session went wrong, quoted from the transcript), patterns (behaviors that recurred across multiple sessions rather than one-off mistakes), and rule proposals (ready-to-copy additions to your project's AGENT.md or new skill files to prevent the same friction in future). All output goes into a .learn folder inside your project, containing a run summary, a friction log with direct quotes, a patterns analysis file, and proposed skill or documentation files. A stamp file tracks when the skill last ran so subsequent runs only process newer sessions. Deleting it forces a full reprocess. It pairs with companion tools called SkillsMCP and claude-mem for skill management and long-term memory. Most useful for developers who work heavily in AI-assisted coding workflows and want past sessions to automatically surface better rules and habits.