Skip to content
Get packs

Guide · Agent skill

How to use the autoskill agent skill: propose skills from your work history

Updated · written with AI from the project's own documentation and checked against our catalogue data

Autoskill is an agent skill from the K-Dense-AI scientific-agent-skills repository. When you explicitly ask, it reads your recent activity from a local Screenpipe daemon, groups repeated workflows, checks them against existing scientific skills, and writes staged draft skills for you to review. This guide covers what it does, how to install it, how to trigger it, and what to configure first so you know what it will read.

Autoskill: Propose Skills From Your Work HistoryK-Dense-AI/scientific-agent-skills · MIT

Quick start

  1. 1Download and unzip. Download the autoskill ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Start Screenpipe. Run the Screenpipe daemon so its HTTP API listens on localhost:3030, with the privacy exclusions you chose.
  3. 3Set the token. If authentication is enabled, export SCREENPIPE_TOKEN so protected search routes accept requests.
  4. 4Prepare Python. Install httpx, PyYAML, and sentence-transformers in a separate environment with Python 3.10 or newer.
  5. 5Load a local model. Start the LM Studio local server and load a chat model with a stable ID, then set local.model in the config.
  6. 6Ask for an analysis. Request an analysis of a time window and ask for proposed skills, starting with a dry run if you want to review first.
  7. 7Review and promote. Read the staged drafts in the proposal folder, edit them, and promote only the ones you want.

What does the autoskill skill make the agent do?

Autoskill turns your own workflow history into candidate skills. It queries the Screenpipe HTTP API on localhost for a time window you name, clusters repeated patterns of apps and window titles, and compares each cluster against the scientific skills already in the repository.

The output is a staged folder of draft skills or composition recipes. Nothing is installed automatically: you review, edit, and promote the drafts yourself with the included promote script.

  • Reads a history window you specify, such as the last 4 hours, day, or week
  • Detects repeated research workflows
  • Matches them against existing scientific skills
  • Writes staged drafts or composition recipes for review

Install the autoskill skill from the ZIP

Download the autoskill ZIP from this directory page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

  • Personal: ~/.claude/skills/autoskill
  • Project: .claude/skills/autoskill inside your project
~/.claude/skills/

Prepare Screenpipe, the token and Python

Autoskill needs a running Screenpipe daemon, normally at http://localhost:3030. Protected routes such as search require a bearer token even on loopback, so if your instance has auth enabled, export the token before running the skill. Health checks do not validate search access.

Create a separate Python environment with Python 3.10 or newer and the packages httpx, PyYAML, and sentence-transformers. The source's recommended setup uses Python 3.12 in a dedicated folder. The embedding model downloads on first use, so you need network access for that step.

  • Screenpipe daemon running on localhost:3030
  • SCREENPIPE_TOKEN set if authentication is enabled
  • Python 3.10+ with httpx, PyYAML, and sentence-transformers
  • LM Studio local server with a chat model loaded (default LLM path)
export SCREENPIPE_TOKEN="$(screenpipe auth token)"

Review capture exclusions before collecting history

Screenpipe captures passively, so the skill can see whatever it recorded. The skill's reference checklist lists app and window-title substrings to exclude, such as password managers and private browsing windows. Apply them in Screenpipe settings or as repeated --ignored-windows arguments when you start recording.

Filtering cannot remove sensitive material already captured, and the redaction step is a best-effort scrubber. Check your exclusions with synthetic windows before relying on them.

  • Exclusions are literal, case-insensitive substrings, not globs
  • Dry-run mode writes a plan.md you can review before any LLM step
  • Configure capture filtering before you collect history

Trigger the skill with a natural request

Ask for an analysis of a specific window and ask for proposals. The skill is built for explicit, on-demand requests and should not run in the background. Do not use it for general questions about Screenpipe or for real-time screen queries.

If you want to look at the clustered workflows without any LLM call, ask for a dry run.

  • "Analyze my last 4 hours and propose new skills."
  • "Look at what I've been doing and tell me what's not covered yet."
  • "Draft a skill from my recent workflow."
  • "Find composition recipes for workflows I repeat."

Safety, network use and the LLM backend

The default LLM backend is a local LM Studio server on localhost:1234. Summaries stay on your machine while the endpoint is loopback. Cloud backends, Claude or Foundry, are opt-in and receive redacted cluster summaries and matched skill descriptions. Detection and embedding inference run locally, but the embedding model download does use the network.

The skill runs scripts and makes authenticated HTTP requests. Redaction removes recognizable emails, API keys, bearer tokens, and selected phone formats, but it does not guarantee that unpublished research details or window titles are removed. Review the plan before choosing a cloud backend.

  • Default: local LM Studio, summaries stay on the machine
  • Opt-in: Claude (ANTHROPIC_API_KEY) or Foundry (FOUNDRY_API_KEY)
  • Optional env vars: SCREENPIPE_TOKEN, LM_API_TOKEN
  • Screenpipe connections allow plain HTTP only on loopback

Tips for better results

Use a narrow time window first so the clusters are easy to review. Read the draft skills before promoting them, since they are proposals grounded in your usage rather than verified workflows. Treat similarity scores as a retrieval heuristic, not proof that two workflows are equivalent.

If the embedding model is already cached, set embeddings.local_files_only to true so no download is attempted. Long summaries may lose detail because the embedding model truncates inputs beyond 256 word pieces.

  • Start with a short window such as the last 4 hours
  • Use dry-run first to inspect plan.md
  • Set embeddings.local_files_only: true for an existing cache

What people use it for

  • Finding repeated literature-review or data-cleaning steps that could become a reusable skill
  • Checking which of your recent analysis workflows are not yet covered by existing scientific skills
  • Drafting a composition recipe that chains existing skills for a workflow you repeat each week
  • Reviewing a dry-run plan.md to see how your work history clusters before deciding anything

Frequently asked questions

How do I install the autoskill skill?

Download the autoskill ZIP from this page and unzip it into ~/.claude/skills/ for personal use or a project's .claude/skills/ folder. In the Claude apps you can also upload it under Settings → Capabilities → Skills.

What does autoskill require?

It requires Python 3.10 or newer with httpx, PyYAML, and sentence-transformers, plus a running Screenpipe daemon at localhost:3030. It also needs an LLM backend: LM Studio's local server by default, or an opt-in Claude or Foundry account. Network access is needed for the first embedding-model download.

Is autoskill free and what is its licence?

The skill is published under the MIT licence. The source does not state pricing for the skill itself, and the Screenpipe and cloud LLM services are separate tools with their own terms.

Does autoskill send my screen data to the cloud?

Not by default. The default backend is a local LM Studio server, and summaries stay on your machine while that endpoint is loopback. Cloud backends are opt-in, and they receive redacted app and title cluster summaries and matched skill descriptions, not raw OCR.

What is the difference between autoskill and Screenpipe?

Screenpipe is a tool that passively records which apps and window titles you use. Autoskill reads that history through Screenpipe's local API on demand and turns repeated workflows into skill drafts. It does not replace Screenpipe and is not a general Screenpipe question-answering tool.

Why does autoskill say it cannot connect to Screenpipe?

The skill raises ScreenpipeUnreachable when the daemon is not running. Start Screenpipe so its HTTP API listens on localhost:3030. If auth is enabled, also export SCREENPIPE_TOKEN, since search routes require a bearer token even on loopback.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Autoskill suits researchers and skill-library maintainers who already run Screenpipe locally, understand what it records, and want reviewable suggestions based on real usage. It is a poor fit if you have not reviewed capture exclusions, handle confidential material, or want general Screenpipe help, since redaction is only a best-effort scrubber.

Download the skill