autoskill

SUMMARY

What it does

autoskill is an Agent Skill that analyzes a user's recent on-screen activity (captured by the local screenpipe daemon) to identify repeated research workflows, matches them against the existing 135 skills in the scientific-agent-skills repository, and drafts new skills or composition recipes for uncovered patterns. It requires a locally running screenpipe daemon on localhost:3030 with a valid SCREENPIPE_TOKEN; without it, the skill raises ScreenpipeUnreachable. The skill performs all detection and clustering locally, redacts sensitive data (emails, API keys, tokens, phone numbers) before any LLM call, and defaults to a local LLM backend (LM Studio) to keep data on-machine. Cloud backends (Claude, Foundry) are opt-in. The skill is invoked on-demand by the user with a time window, and outputs a staged folder of proposals (report.md, new-skills/, composition-recipes/) for user review and promotion. It includes a doctor subcommand for preflight checks, a run subcommand for the pipeline, and a promote subcommand to move approved proposals into the skills directory. The skill is part of the Scientific Agent Skills library by K-Dense, licensed under MIT, and cites a related arXiv paper when used in research outputs.