coach-checkin

name

coach-checkin

description

Run the opt-in Lifestyles of the AI-Native workshop check-in. Use at the start of the workshop and again at the close. Auto-detects opening vs closing, runs the short interview, records your AI-Native score from a LOCAL scan at the open, and at the close posts your answers and your before→after score (anonymously) to the live room board. Triggers on "check in", "coach check-in", "workshop check-in", "run my workshop check-in", "closing check-in", or "share my workflow".

Coach Check-in

A short, opt-in check-in for Lifestyles of the AI-Native workshop. You run it twice — once walking in, once at the close. Each run does two things: a quick interview (what's eating your week → what you built) and an AI-Native score read from your local setup. Together they feed the live room board: where the room's toil is, which hooks and scheduled tasks to build, the engineering-hours/week the room reclaims, and the room's before→after AI-Native score.

Privacy is the whole point. The score is computed by a local scan of your Claude setup (do you have hooks? a CLAUDE.md? worktrees? scheduled jobs?) that never leaves your machine — only the derived score numbers and the answers you explicitly confirm are sent. No file contents, no git log, no transcripts are ever transmitted. Responses are anonymous: a random participant id plus your role/stack — no name, no email.

How it works

The skill auto-detects which check-in to run using a marker file (.aie-coach-state.json) in the repo root: no marker means this is the opening (pre) check-in; an existing marker means the closing (post) check-in. The same anonymous participant id is reused across both runs so the board can connect your before and after. All file state, payload assembly, and network I/O live in scripts/submit.ts — do not write the marker, build the JSON, or POST by hand.

Run every command from the workshop repo root (so the marker is written there, where detection looks for it).

Procedure

1. Detect the phase. Run:

   bun skills/coach-checkin/scripts/submit.ts detect
   

   Read the phase field from its JSON output ("pre" or "post").

2. Run the matching interview (fixed prompts; ask one brief follow-up only if an answer is empty or evasive):

   • If phase is pre:
     1. Ask their role / stack as free text (e.g. "What's your role and main stack?"). This is the role field — send it verbatim; the board classifies it.
     2. Ask: "What dev task eats the most of your week?" → time_sink.
     3. Ask: "What's the most repetitive thing you still do by hand?" → friction.
     4. Ask: "What would you most love to automate or speed up today?" → goal.
   • If phase is post: (the role is already known — do not ask again)
     1. Ask: "What did you wire up today — a hook, a skill, a scheduled task?" → built.
     2. Ask: "What are you going to automate next?" → next.

3. Show an editable summary. Display the role (pre only) and each answer back to the participant, then ask with AskUserQuestion: Send, Edit, or Cancel.

   • Edit: revise the named answer and re-show the summary.
   • Cancel: stop. Nothing is written or sent.

4. Send on confirmation. Pipe the collected data as JSON to the submit command. The answers object is keyed by the exact question keys above. Include role only for the pre run.

   echo '{"role":"Backend / Go","answers":{"time_sink":"...","friction":"...","goal":"..."},"confirmed":true}' \
     | bun skills/coach-checkin/scripts/submit.ts submit
   

   For the post run, omit role (it comes from the marker):

   echo '{"answers":{"built":"...","next":"..."},"confirmed":true}' \
     | bun skills/coach-checkin/scripts/submit.ts submit
   

5. Record / send the AI-Native score using the aie-coach MCP tools (loaded when the repo is trusted). The score is read from a local scan — see the privacy note above.

   • Opening (pre): after the answers are sent, call the coach_scan tool (no args). This records the opening baseline. Tell the participant their starting score out of 100 and the weakest pillar, framed as an invitation: "That's your starting line — pick one of these to actually try today."
   • Closing (post): after the answers are sent, confirm consent (the same Send confirmation covers it) and call coach_submit_score with { "confirmed": true }. This sends the before→after to the board. Report the delta: "You went from X → Y — that's now on the board."
     • If it returns {"sent":false,"reason":"no_baseline"}, the opening scan never ran — call coach_scan once, then retry coach_submit_score.

6. Report the result plainly from the command's JSON output:

   • {"sent":true,...} → "Sent — thanks, your pulse is on the board."
   • {"sent":false,"outbox":"..."} → "Saved locally (the board was unreachable) — a facilitator can flush it later." Your pre/post state is still recorded.

Notes

• All commands run from the workshop repo root; the marker (.aie-coach-state.json) and any outbox files land there.
• The board URL and auth token are baked into scripts/submit.ts; set the WORKER_URL / WORKER_TOKEN environment variables to override them (the facilitator sets these once the board is deployed).
• The scripts live in the repo's skills/coach-checkin/ (this SKILL.md lives in .claude/skills/ so Claude Code auto-loads it; commands run from the repo root, so the bun skills/coach-checkin/... paths resolve). The payload contract is in skills/coach-checkin/scripts/feedback-contract.schema.json, described in CONTRACT.md.
• Never send anything the participant didn't confirm. The AI-Native score is read by a local scan (coach_scan) of their Claude setup, but only the derived score numbers leave the machine — never file contents, git log, or transcripts. Interview answers are volunteered-only.