autoskill — Mine New Skills From Your Own Workflow

What it does

autoskill reads the screen history captured passively by a locally running screenpipe daemon and turns your repeated workflows into skill proposals.

• Fetch — paginates screenpipe's /search API on localhost:3030 for the time window you request.
• Redact — regex-scrubs emails, API keys, bearer tokens and phone numbers as defense-in-depth.
• Cluster — segments sessions on idle gaps (default 10 min) and groups them by app signature.
• Match — ranks each cluster against the repo's 135 existing skills using local sentence-transformers embeddings.
• Synthesize — an LLM judge labels each cluster reuse, compose, or novel and writes a SKILL.md body.

Output lands in ~/.autoskill/proposed/<timestamp>/ (report.md, new-skills/, composition-recipes/) for human review, then promote moves an approved draft into your real skills directory. The default backend is a local LM Studio model, so nothing leaves the machine; Claude and Foundry backends are strictly opt-in.

Who it's for

• Researchers who repeat the same multi-app routine weekly and want evidence-based automation candidates.
• Skill authors who need real usage data to decide what to build next.
• Privacy-conscious users who want detection, embeddings, and drafting to run entirely on-device.

Examples

1. "Analyze my last 4 hours and propose new skills" → run the pipeline, open report.md, and get a zotero-pubmed-helper draft for your Zotero → PubMed → notes loop.
2. Use --dry-run to emit only plan.md so you can inspect exactly which timeline will be analyzed before any LLM call.
3. When a cluster scores high against literature-review + citation-management, you get a thin composition recipe that chains both skills in order — no new skill needed.

Install the skill I found on Claude Skill Mart.
Copy the skills/autoskill folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/autoskill/.
When it's done, tell me in one line what this skill can do.