setup

name	setup
description	First-run initialization of an earbrief instance — interview the owner, write config.md, personalize sources/curriculum, publish the player artifact, create the cloud routines. Run when config.md is missing.

earbrief /setup

You are initializing a fresh earbrief instance (a repo created from the earbrief template). The outcome: a written config.md, personalized sources.md/curriculum.md, a published player artifact, two scheduled cloud routines, everything committed and pushed.

0. Guards

• If config.md exists, this instance is already initialized. Show its contents and ask what to change instead of re-running the flow (a re-run would mint a duplicate artifact and duplicate routines).
• Check git remote -v. If there is no origin, stop and tell the user to create their GitHub repo (private recommended — the repo will accumulate their personal listening log) and push first.
• Warn if the repo is public: gh repo view --json visibility (skip silently if gh is unavailable).

1. Interview

Ask (AskUserQuestion where it fits, free text otherwise):

1. Timezone and schedule — daily digest time (default 07:00 local) and weekly deep-dive day+time (default Saturday 08:00 local). Convert to UTC for scheduling.
2. Secondary language — Spanish or none. English is always canonical. Only Spanish is supported as secondary in this version (the player's voice mapping and AUTO mode are EN/ES); other languages require editing player/template.html.
3. Listener profile — who they are, what they already know, what depth they want. One short paragraph; this steers every episode's tone and level.
4. Domain — keep the default AI/engineering sources and curriculum, or describe their field. If they change domain, rewrite sources.md (research real, currently-active sources for that field; keep the tier structure and the x-prefix disable convention and the editorial rules section) and curriculum.md (build tracks from their stated gaps; keep the first-unchecked convention and Track E).
5. Episode length — daily words (default 2200 ≈ 15 min) and deep-dive words (default 3000 ≈ 20 min).

2. Write config.md

# Config — instance settings

Written by /setup. Routines, CLAUDE.md, and chat ops read instance values from here.

- player_artifact_url: (pending first publish)
- daily_routine_id: (pending)
- weekly_routine_id: (pending)
- timezone: <IANA tz>
- daily_schedule: <HH:MM local / HH:MM UTC>
- weekly_schedule: <Day HH:MM local / HH:MM UTC>
- secondary_language: es | none
- daily_words: 2200
- deepdive_words: 3000

## Listener profile

<the paragraph from the interview>

Also update the profile line in curriculum.md with the same paragraph (condensed to one line).

3. Build and publish the player

1. python3 player/build.py — with zero episodes it prints built player/player.html with 0 episodes; that is correct for a fresh instance.
2. Publish with the Artifact tool: file_path player/player.html, no url param (first publish mints this instance's own address), favicon 📻, description 'Audio player for the daily AI briefing'.
3. Record the returned URL as player_artifact_url in config.md. Every later publish MUST pass this URL as url or the phone bookmark breaks.

4. Create the routines

Use the schedule skill (scheduled cloud agents). The user's repo must be connected to Claude's GitHub app — if scheduling fails on repo access, send them to claude.ai/code to connect the repo, then retry.

• Daily digest — cron from the interview's daily time (UTC). Prompt: You are a scheduled routine of the earbrief pipeline. The repo is already cloned in your working directory. Read routines/daily.md in this repo and follow it exactly. Suggested model: sonnet.
• Weekly deep-dive — cron from the weekly day+time (UTC). Same prompt with routines/weekly.md. Suggested model: opus.

Both routines need allowed tools beyond the defaults: Bash, Read, Write, Edit, Glob, Grep, WebSearch, WebFetch, Artifact (WebSearch/WebFetch for research, Artifact to republish the player).

Record both trigger IDs in config.md. If routine creation is unavailable in this environment, write (create manually) in config.md and print step-by-step instructions for claude.ai/code/routines instead — including both prompts verbatim.

5. Upstream remote

If origin is not the template repo itself, add the template as upstream for future updates (skip if it already exists):

git remote add upstream https://github.com/mblasi/earbrief.git

6. Commit and hand over

1. Commit everything: initialize earbrief instance, push.
2. Print, in this order: the player URL with "bookmark this on your phone"; when the first episode will appear (next daily run, or offer to trigger the daily routine once now); the chat-ops cheat sheet (mark listened, add/disable source, promote topic, rebuild player, update from upstream — see CLAUDE.md).