Autoskill

---
name: autoskill
description: Observe the user's screen via screenpipe, detect repeated research workflows, match them against existing research-agent-skills, and draft new skills (or composition recipes that chain existing ones) for the patterns not yet covered. Use when the user asks to analyze their recent work and propose skills based on what they actually do. Requires the screenpipe daemon (https://github.com/screenpipe/screenpipe) running locally on port 3030 — the skill has no other data source and will refuse to run if screenpipe is unreachable. All detection runs locally; only redacted cluster summaries reach the LLM.
license: MIT
allowed-tools: Read Write Edit Bash
metadata:
  version: '1.4'
  category: research-automation
  maintainer: Kalaris Labs
  openclaw:
    requires:
      bins:
      - screenpipe
    primaryEnv: SCREENPIPE_TOKEN
    envVars:
    - name: SCREENPIPE_TOKEN
      required: true
      description: Auth token for the local screenpipe daemon.
    - name: ANTHROPIC_API_KEY
      required: false
      description: For Claude API calls during skill drafting.
    - name: FOUNDRY_API_KEY
      required: false
      description: Optional Foundry access for drafting.
---

# autoskill

> **Requires a running [screenpipe](https://github.com/screenpipe/screenpipe) daemon.** This skill has no alternate data source — it reads exclusively from the local screenpipe HTTP API (default `http://localhost:3030`). If the daemon isn't running, `run()` raises `ScreenpipeUnreachable` with install instructions.

> **Network access & environment variables.** This skill makes authenticated HTTP requests to (a) the user's local screenpipe daemon on loopback, and (b) the user-configured LLM backend — one of `http://localhost:1234/v1` (LM Studio, default), `https://api.anthropic.com` (opt-in Claude), or a user-supplied BYOK Foundry gateway. The skill reads three environment variables — `SCREENPIPE_TOKEN`, `ANTHROPIC_API_KEY`, `FOUNDRY_API_KEY` — and uses each only to authenticate to the single endpoint its name implies. No other network destinations, no telemetry, no data egress to any third party.

## Overview

Turn the user's own workflow history — captured passively by the local [screenpipe](https://github.com/screenpipe/screenpipe) daemon — into new skills. This skill is on-demand: the user invokes it with a time window, it queries screenpipe's local HTTP API, clusters repeated workflow patterns, compares each pattern against the existing skills in this repo, and produces a staged folder of proposals the user can review, edit, and promote.

## When to Use This Skill

Invoke this skill when the user asks to:
- "Analyze my last 4 hours / day / week 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."

Do **not** invoke it for one-off questions about screenpipe itself, for real-time screen queries, or without an explicit user request — the skill analyzes sensitive local content and must stay explicitly user-triggered.

## Privacy Posture

- **Screenpipe handles app/window filtering at capture time.** Install a starter deny-list by copying `references/screenpipe-config.yaml` into the user's screenpipe config. Sensitive apps (password managers, messaging, banking) are never OCR'd in the first place.
- **Raw OCR never leaves the machine.** `scripts/fetch_window.py` pulls data over localhost HTTP. `scripts/cluster.py` reduces the timeline to app/duration/title summaries. `scripts/redact.py` strips emails, API keys, bearer tokens, and phone numbers as defense-in-depth before any cluster summary reaches the LLM.
- **LLM backend defaults to `local`.** The recommended setup is [LM Studio](https://lmstudio.ai/) running `Gemma-4-31B-it` — strong reasoning at a size that fits on most workstation GPUs, and no data ever leaves your machine. Cloud backends (`claude`, `foundry`) are opt-in and documented in `config.yaml` for users who explicitly want them. Detection and embeddings always run locally regardless of backend choice.
- **Dry-run mode** (`--plan`) prints the exact timeline that will be analyzed before any LLM call.
- **TLS for localhost** (optional, for corporate policy): see `references/https-proxy.md` for the Caddy pattern.

## Prerequisites

### 1. Screenpipe daemon

Either install the official release or build from source. Either way the daemon binds HTTP on `localhost:3030` by default.

**From source** (recommended if you want the CLI daemon without the desktop GUI):

```bash
git clone --depth 1 https://github.com/mediar-ai/screenpipe.git
cd screenpipe
cargo build -p screenpipe-engine --release
# System deps (macOS): cmake + full Xcode.app (not just Command Line Tools).
#   brew install cmake
#   # if xcodebuild plug-ins error: sudo xcodebuild -runFirstLaunch
./target/release/screenpipe doctor   # confirm permissions + ffmpeg
./target/release/screenpipe record --disable-audio --use-pii-removal
```

First run will prompt for macOS Screen Recording permission. Grant it and relaunch.

### 2. Screenpipe API token

The local API now requires bearer auth. Retrieve your token and export it:

```bash
export SCREENPIPE_TOKEN=$(screenpipe auth token)
```

(Or set `screenpipe.token` directly in `config.yaml` — env var is preferred since it keeps secrets out of version control.)

### 3. Python environment

Via `pipenv` from the repo root:

```bash
pipenv install httpx pyyaml sentence-transformers
```

The embedding model (`sentence-transformers/all-MiniLM-L6-v2`, ~80 MB) downloads on first run.

### 4. Local LLM (default path) — LM Studio

- Install [LM Studio](https://lmstudio.ai/).
- Download `Gemma-4-31B-it` (or another strong reasoning model; adjust `local.model` in `config.yaml`).
- Load it via the CLI for headless use (no GUI required):

```bash
lms load gemma-4-31b-it --context-length 131072 --gpu max -y
lms status   # confirm server running on :1234
```

### 5. Cloud LLM backends (optional, opt-in)

Only if you explicitly opt out of local:
- `claude`: set `ANTHROPIC_API_KEY`, flip `backend: claude` in `config.yaml`.
- `foundry`: set `FOUNDRY_API_KEY`, flip `backend: foundry`, set `foundry.endpoint` to your corporate gateway URL.

## Architecture

```
screenpipe daemon (user-installed)
        │  HTTP on localhost:3030
        ▼
scripts/fetch_window.py    → normalized timeline events
scripts/redact.py          → regex scrub (defense-in-depth)
scripts/cluster.py         → sessions + clusters (local only)
scripts/match_skills.py    → top-k vs existing 135 skills (local embeddings)
scripts/synthesize.py      → LLM judge: reuse / compose / novel
        │
        ▼
~/.autoskill/proposed/<timestamp>/        (default; override with --out)
  ├── report.md
  ├── composition-recipes/<name>/SKILL.md
  └── new-skills/<name>/SKILL.md

scripts/promote.py         → user-approved proposal → skills/<name>/
```

## Workflow

The skill ships a unified CLI at `scripts/autoskill.py` with three subcommands:

```bash
python scripts/autoskill.py doctor   --config config.yaml --skills-dir ../
python scripts/autoskill.py run      --start ... --end ... --config config.yaml
python scripts/autoskill.py promote  --proposed ~/.autoskill/proposed/<ts> --skills-dir ../ --name <skill>
```

### 0. Preflight with `doctor`

Before a full run, verify every dependency in one shot:

```bash
python scripts/autoskill.py doctor \
  --config skills/autoskill/config.yaml \
  --skills-dir skills
```

The report covers `config` (backend choice valid), `skills_dir` (exists), `screenpipe` (reachable + authed), and `llm` (LM Studio serving or API key present). Non-zero exit on any failure, with the offending line marked `error`.

### 1. Run the pipeline

```bash
export SCREENPIPE_TOKEN=$(screenpipe auth token)
python scripts/autoskill.py run \
  --start "2026-04-17T00:00:00Z" \
  --end   "2026-04-17T23:59:59Z" \
  --config skills/autoskill/config.yaml \
  --skills-dir skills
```

Proposals land in `~/.autoskill/proposed/<timestamp>/` by default, keeping experimental output out of the skills repo. Pass `--out PATH` to override.

Internally:
1. **Fetch** — `fetch_window` paginates screenpipe's `/search` endpoint, normalizes events to `{ts, app, window_title, text, content_type}`.
2. **Redact** — `redact` scrubs emails, API keys, bearer tokens, phones from OCR text and window titles as defense-in-depth over screenpipe's own PII removal.
3. **Cluster** — `segment_sessions` splits on idle gaps (default 10 min) and drops short sessions; `cluster_sessions` groups sessions by app-signature and keeps clusters of size `min_cluster_size` (default 2).
4. **Match** — `load_skill_descriptions` reads frontmatter from every `SKILL.md` in `skills/`; `top_k_matches` ranks each cluster against all skills using local `sentence-transformers` embeddings (cosine similarity).
5. **Synthesize** — `synthesize` prompts the configured LLM backend to classify each cluster as `reuse`, `compose`, or `novel` and emit a SKILL.md body where appropriate.
6. **Report** — writes `<out_dir>/<ts>/report.md`, plus `new-skills/<name>/SKILL.md` or `composition-recipes/<name>/SKILL.md` for each proposal.

Add `--dry-run` to stop after clustering; this skips the LLM (and the sentence-transformers load), writing only `plan.md` for inspection.

### 2. Review and promote

Open `~/.autoskill/proposed/<ts>/report.md`, edit drafts in place, delete anything you don't want. Then:

```bash
python scripts/autoskill.py promote \
  --proposed ~/.autoskill/proposed/2026-04-17T14-30-00 \
  --skills-dir skills \
  --name zotero-pubmed-helper
```

`promote` moves the directory into `skills/<name>/`, refusing to overwrite an existing skill. Exits non-zero with a friendly error if the proposal isn't found or the target already exists.

## Configuration

See `config.yaml` for the full shape. Default values (local-first):

```yaml
backend: local
local:
  endpoint: http://localhost:1234/v1   # LM Studio's Developer server
  model: Gemma-4-31B-it

screenpipe:
  url: http://localhost:3030           # or https://screenpipe.local via Caddy

cluster:
  min_session_minutes: 5
  idle_gap_minutes: 10
  min_cluster_size: 2
```

To opt into a cloud backend:

```yaml
backend: claude                         # or foundry
claude:
  model: claude-opus-4-7
```

## Composition recipes vs new skills

- **compose**: the LLM judged that chaining existing skills covers the workflow. The emitted SKILL.md is intentionally thin — frontmatter + a "Workflow" section that invokes existing skills in order. The same agent runtime that discovered the skill can then invoke it end-to-end.
- **novel**: no combination of existing skills covers it. A fuller SKILL.md is drafted, still following repo conventions (frontmatter, Overview, When to Use, Workflow). The user should always review new-skill drafts before promoting.

## Testing

The skill is covered by a small pytest suite at `tests/autoskill/` in the repository root. Each script is unit-tested in isolation with dependency injection (mock HTTP transport, stub backend, stub embedder):

```bash
python -m pytest tests/autoskill -v
```

## Composition with other skills in this repo

The autoskill's embedding index covers all 135 sibling skills. Workflows that look like scientific writing will match `scientific-writing` / `literature-review` / `citation-management`; figure work will match `scientific-schematics` / `generate-image` / `infographics`; slide prep matches `scientific-slides` / `pptx`; etc. When a cluster scores high against two or three sibling skills the emitted composition recipe names them explicitly, so the user's future agent invocations use the optimized paths already documented in this repo.

## Agent operating procedure

1. **Check the environment.** Confirm the objective, evaluation metric, budget (time, compute, money) and stopping criteria.
2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
3. **Run a small version first.** Run one iteration end to end and inspect logs before scaling up.
4. **Execute the full task** using the instructions and references above.
5. **Validate the result.** Evaluate on held-out data; keep an audit trail of every iteration, hypothesis and result.
6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.

| If this happens | Do this |
|---|---|
| An iteration produces suspiciously large gains | Check for leakage or evaluation bugs before accepting the result. |
| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |
| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |

**Integrity rules**

- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
- Record negative results; never select or report only favourable runs.
- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

## Related skills

- `research-skill-creator`: Create, improve and test agent skills for research workflows (paper writing, lab protocols, analysis pipelines, domain databases) that meet…