loom

Files (1)

• SKILL.md

SKILL.md (first lines)

# Loom pipeline

This plugin turns the signed-in user's Loom recordings into reviewed action plans and publishes
them to Notion (and optionally Google Drive). It runs entirely on the user's Claude subscription
(no LLM API keys): Claude itself performs the analysis. Loom has no public API, so it drives an
authenticated local browser session - this only works on the user's own machine.

## Hard constraints
- Local only. The Loom session lives in a local browser storage state; nothing here can run on a
  remote/hosted agent.
- No LLM API keys. The analysis step is done by you (the Claude session), not an SDK call.

## State location
All state lives in a stable per-user data dir (default `~/.loom-daddy`, override with
`LOOM_DADDY_DATA_DIR`): the Loom session, the processed-ledger/watermark, artifacts, and a
small `config.json` (cached publish destination). It is independent of the plugin's versioned
cache dir, so it survives plugin upgrades. On startup, login/setup migrate an existing session
from the old layout automatically.

## How the pieces fit
Scripts (no LLM) do the IO; you do the analysis across an artifact boundary in `artifacts/`:

```
download (script)  ->  analyze (YOU)        ->  publish (Notion MCP + Drive script)
<id>.transcript.json   <id>.plan.json           Notion page + optional Drive copy
<id>.meta.json
```

All scripts run from the plugin root with the plugin's own dependencies and `.env`:
```bash
cd "${CLAUDE_PLUGIN_ROOT}" && npx tsx --env-file-if-exists=.env scripts/<name>.ts <args>
```

## Commands (explicit)
- `/loom-daddy:setup` - check session/data-dir/Notion-MCP/optional-Drive, guide first-time setup
- `/loom-daddy:loom-auth` - mint/refresh the Loom session (opens a browser; saves automatically)
- `/loom-daddy:status` - check whether the session is valid
- `/loom-daddy:list` - list recent recordings (new vs already processed)
- `/loom-daddy:analyze <id>` - fetch a transcript and produce an action plan in chat (read-only)
- `/loom-daddy:render <id>` - render a plan to markdown without publishing
- `/loom-daddy:publish <id> [--to notion|drive|both]` - publish a plan
- `/loom-daddy:process` - run the whole pipeline over all new recordings
- `/loom-daddy:schedule install|remove|status` - manage the local background poller

## State rules (important)
- `analyze` is READ-ONLY: it never marks a recording processed, so the poller can still pick it up.
- `publish` marks the recording processed (ledger) on a successful commit, so the poller will not
  republish a hand-published recording.
- `process` marks each recording after publishing and advances the time watermark once per run.

## Publishing
- Notion uses the connected Notion MCP in BOTH the interactive and the scheduled paths (a headless
  `claude -p` run can reach the claude.ai Notion connector - verified). No `NOTION_TOKEN`. The
  target database is resolved from a name you give once, then cached in `config.json`. The page
  mirrors the built-in layout: title = recording name; Summary; Decisions (timestamp links);
  Action Items (table: item, owner, priority, timeframe, source link); Open Questions.
- Drive is OPTIONAL and uses the built-in publisher (service account). It is skipped silently when
  `DRIVE_FOLDER_ID` / `GOOGLE_APPLICATION_CREDENTIALS` are unset.

## Setup
Run `/loom-daddy:setup`. The only hard requirement is a valid Loom session
(`/loom-daddy:loom-auth`) plus a connected Notion MCP for publishing. No LLM API key. Optional env
(in `${CLAUDE_PLUGIN_ROOT}/.env`): `DRIVE_FOLDER_ID` + `GOOGLE_APPLICATION_CREDENTIALS` to enable
Drive; `LOOM_DADDY_DATA_DIR` to relocate state; `LOGIN_TIMEOUT_MS` to extend the login wait.

From fontezbrooks/loomDaddy, MIT.

More from fontezbrooks/loomDaddy