@claude comments)Each person on the board sets up their own routine, one for each board repo. A routine acts for one person: it uses their Claude usage and limits, and its work is recorded against them. Typing @claude ... in a task comment on the web board then starts your routine for that board on that task, after you confirm “Send to Claude?”.
This guide is for the comment-triggered board assistant. Keep its connectors disabled. For a scheduled import from
Google Drive or another meeting-notes source, create a separate least-privilege routine and follow
.claude/skills/keeptrack-notes/references/routines.md. That guide includes a copyable prompt and an incremental log
of source document revisions and resulting task numbers, without storing transcript text. It defaults to a rolling
seven-day lookback, skips documents already logged at the same revision, and matches repeated actions across meetings.
It also explains how to recover a verified legacy import and how explicit versus uncertain completion evidence changes
an existing task.
If you would like an agent to walk you through this, ask Claude Code in this repo: “set up my board routine”. It uses the board-routine-setup skill, checks each step, and tells you when it needs you to paste a secret. An agent can guide, check and verify, but it cannot create the API trigger or paste secrets for you (see “What only you can do”).
| Item | Each person? | Where it lives |
|---|---|---|
| A routine (their own) | Yes | claude.ai (their account) |
| Routine trigger URL + token | Yes, one per routine | Their browser (board Settings → Agents & integrations); also held by cron-job.org for the minute or two a send takes |
BOARD_TOKEN (GitHub token, Contents read/write on this repo) |
Yes, their own | The routine’s cloud environment on claude.ai (never in the repo) |
| cron-job.org API key | One per cron-job.org account. Each person can have their own free account (simplest), or share one key | Their browser (board Settings → Agents & integrations) |
| GitHub token for the web board | Yes (already done) | Their browser (board Settings → Connection) |
Nothing secret is ever committed to this repo. “Copy settings code” on the board moves your browser’s values to your phone or another browser (tick “Include tokens and keys” to take the tokens too).
You can ask Claude (Claude Code in this repo, the Claude desktop app, or claude.ai) to do the setup. In the board, open Settings → Agents & integrations → @claude in task comments, switch Use @claude on, then Let Claude guide you instead → Copy setup prompt and paste it into a new chat. It points Claude at this guide and offers two modes: guide me step by step, or do it in my browser (Claude in Chrome or the built-in browser: you stay signed in, Claude does the clicking and the non-secret fields).
Whichever mode, Claude must never type, read back or store a secret. At each secret it stops, tells you where to click and what to paste, and waits. The secrets are the routine’s trigger token, BOARD_TOKEN, and the cron-job.org key.
The same prompt, for pasting into any chat (replace the placeholders):
Please set up my Claude routine for the task board in <owner>/<repo>, so that typing @claude in a task comment starts it.
My GitHub username is <your-github-username>. The board repo is <owner>/<repo>.
Read these first:
- https://github.com/<owner>/<repo>/blob/master/board/ROUTINE-SETUP.md (the runbook)
- https://github.com/<owner>/<repo>/blob/master/board/routine-loader.txt (the short prompt to paste into the routine; it points at routine-prompt.md in the repo)
- https://github.com/<owner>/<repo>/blob/master/.claude/skills/board-routine-setup/SKILL.md (what you may and may not do)
Then ask me which mode I want:
A) Guide me: walk me through each step in order and check each one.
B) Do it for me in my browser: if you have a browser tool (Claude in Chrome or the built-in browser), open claude.ai/code/routines, GitHub's fine-grained token page and console.cron-job.org (I am already signed in) and do the clicking and the non-secret fields: routine name, the loader prompt from routine-loader.txt, the repository, a dedicated environment with Trusted network access, no connectors, and the API trigger.
Rules: never type, paste, read back or store a secret (routine trigger token, BOARD_TOKEN, cron-job.org API key). At each secret step, stop, tell me exactly where to click and what to paste, and wait until I say it is done. Remove all connectors from the routine. Finish by running the verification checklist and the first test from the runbook and tell me what passed and failed.
Board assistant (<your name>).board/routine-loader.txt (four lines). It tells the routine to read board/routine-prompt.md from the repo on every run, so you change the routine’s behaviour by editing routine-prompt.md in GitHub, not by editing the routine. The loader must mention the payload, otherwise the routine treats the board’s request text as inert. Edits apply from the next run (the routine reads the default branch, master). Anyone who can write to this repo can change what the routine does, which is already true of the board itself; keep branch protection in mind if the repo audience grows.<owner>/<repo>). The board skill and the heartbeat hook are committed in this repo (.claude/skills/board/, .claude/settings.json), so a routine session picks them up from the clone. No plugin is needed.Use a dedicated environment called Board assistant (<your name>) rather than Default:
BOARD_TOKEN is an environment variable, and environment variables are visible to anyone who uses that environment. A private environment keeps it away from your other routines and sessions.What the routine’s cloud session needs, and what the default Trusted network level gives it:
| Need | How it works | Trusted level enough? |
|---|---|---|
Clone the repo, push claude/ branches, open PRs |
Anthropic’s GitHub proxy, independent of the network level | Yes |
keeptrack.py talking to the board (api.github.com) |
Direct HTTPS from the session | Yes: api.github.com and github.com are on the default allowlist |
Python for keeptrack.py |
Python 3 is pre-installed in the image | Yes |
keeptrack.py saving the board |
The sandbox blocks GitHub API writes, so keeptrack.py saves by committing the changed board files and running git push to master from the clone (allowed). It does this automatically |
Yes |
| Installing packages (pip/npm) | Package registries are on the allowlist | Yes |
| General web research (arbitrary sites) | Not on the allowlist, so shell/script requests to other sites fail with 403 host_not_allowed |
No: use Custom (Trusted list plus the domains you need) or Full |
Recommendation: start with Trusted. It already covers everything the board needs. Switch to Custom or Full only when you want the routine to research the open web. Anthropic’s docs do not say whether Claude’s built-in web search and fetch tools are bound by this setting, so test it: ask the routine to look something up and read the run log. Be aware of the trade-off: with open internet, text a run reads (a board comment, a web page) could trick it into sending data out, and BOARD_TOKEN lives in that environment. If you widen access, keep the token’s scope to this one repository, keep the expiry short, and keep connectors off.
BOARD_TOKENCreate a GitHub fine-grained token: owner rain-ventures-ai, only this repository, Contents: Read and write (add Issues: Read and write if you want issue-driven work), expiry 90 days. Put it in the routine’s environment as BOARD_TOKEN (on Pro/Max use the environment’s API-credentials section so it is not visible to others who use the environment). keeptrack.py uses it to read and write the board files over the GitHub API.
In the routine → Edit → Add another trigger → API. Save. Open the trigger modal, copy the URL (https://api.anthropic.com/v1/claude_code/routines/trig_.../fire) and click Generate token; copy it immediately (shown once). This token can start only this routine.
Browsers cannot call a routine’s URL directly (Anthropic’s endpoint allows no cross-origin calls), but cron-job.org’s API does. Create a free account at https://console.cron-job.org/, then Settings → API → enable → create a key. When you send something to Claude, the board creates a one-off job there that POSTs to your routine, reads the result, then deletes the job. cron-job.org sees your routine token and the task number (never the task’s text) for about one to two minutes.
Under @claude in task comments, switch Use @claude on. The tab shows the same steps as this guide, with a ✓ for each step that is done. It saves each value as you paste it and checks the cron-job.org key at once. Step 5, Send a test to Claude, makes a small test task, sends it to your routine and shows when Claude starts and when it replies.
Test and fix problems (under the steps):
Board request from @<you> for task #1. Run keeptrack.py comment '#1' "routine write test" and report whether it worked. In the run log the comment command must say commented on ... (not write failed: HTTP 403). If it fails, fix BOARD_TOKEN (see Troubleshooting) before going further.@claude add a comment saying hello, press Comment, choose Send to Claude.Comment with @claude → confirm dialog → comment saved with a claim on the card → one-off cron-job.org job (runs at the next whole minute at least ~75 s away) → POST to your routine’s /fire → the page records the session link and deletes the job → the routine claims the task, works, comments, assigns it back to you and marks it done.
| Symptom | Likely cause |
|—|—|
| “To make @claude start your routine, set it up in Settings → Agents & integrations” | One of the three values is missing or the URL is malformed |
| “cron-job.org rejected the API key” | Key typed wrongly or revoked; free accounts allow about 100 API calls a day |
| Card shows “stuck: routine returned HTTP 401” | Routine token wrong or regenerated; generate a new one and update Settings |
| A write says HTTP 403 ... not permitted through this proxy | Claude’s cloud sandbox lets the GitHub API read but blocks its writes. Not a token problem. keeptrack.py switches to saving with git push by itself; if you see this, the routine is running an old keeptrack.py or has BOARD_WRITE=api set |
| A write says git push ... was refused | The sandbox or branch protection would not let the routine push to master. Read the run log for the reason |
| Run starts, reads the board, but every write says HTTP 403 Forbidden (no mention of a proxy) | BOARD_TOKEN can read but not write: its Contents permission is Read-only. Edit the token (github.com/settings/personal-access-tokens → the token → Repository permissions → Contents: Read and write). The token value stays the same, so nothing to re-paste; for an organisation repo an owner may need to approve the change. Then Run now again |
| Reads fail too (404 or 401) | Token missing, expired, owner/repository not selected, or not approved by the organisation |
| Routine started but did nothing useful | The routine’s saved prompt is not the loader, or the loader was edited so it no longer mentions the payload; paste routine-loader.txt again. To change behaviour, edit routine-prompt.md |
| Card claim goes stale after 30 minutes (12 hours for an attended claim) | The routine ended without done; open the session link to see why, then Release claim |
Keep keys in the board’s Settings only. If a key or token is ever pasted into chat or committed, revoke and regenerate it.