Agent Operating Guide
/coach-autopilot
A headless, scheduled pass over your assigned-step queue that drafts every ready step as a suggestion, requests any missing inputs on the steps themselves, and reports the rest in one digest: never committing, never approving.
Work your assigned-step queue unattended. The skill scans your queue with
/coach-workflow-scan, runs each ready step through
/coach-workflow-execute, every one of which drafts a
suggestion for you to review, skips steps already
handled, and emits a single digest of what it drafted, what it needs from you, what’s
blocked, and what’s waiting on your approval. For a step it can’t run because an input is
missing, a document only you hold, a decision only you can make, it posts a one-time
ALL-UPPERCASE INPUT NEEDED suggestion on the step itself, naming exactly what it
needs, then re-checks on every run and drafts the real result once the input arrives. It
runs as you, so it can’t exceed your access, and it never commits or approves
anything: every change waits in your review queue.
When to use
Section titled “When to use”When you want your assigned queue worked automatically on a schedule: ready steps drafted
for you so you approve them rather than author each from scratch. Start with --dry-run to
see what it would draft before it drafts anything. To triage your queue by hand without
acting, use /coach-workflow-scan; to run one step yourself, use
/coach-workflow-execute.
Inputs
Section titled “Inputs”| Flag | Required | Notes |
|---|---|---|
--cap <n> |
No | Max execute dispatches this run: drafts first, then input requests. Default 10; steps beyond the cap are reported as deferred. |
--dry-run |
No | Scan and report what it would draft and would request input for, issuing zero suggestions. The trust-building first pass. |
--limit <n> |
No | Max assigned steps the underlying scan pulls. Default 50. |
Example
Section titled “Example”/coach-autopilot --dry-run scans your queue and lists each ready step under “Would draft”
with what it would produce: writing nothing. Drop the flag and /coach-autopilot drafts
each ready step, up to the cap of 10, then returns one digest: a count line followed by
sections for what it drafted (each with a review-and-approve link), what input it
requested, what still needs you, what’s blocked, what’s awaiting your approval, any
placeholder that was approved without being read, and anything it couldn’t complete.
Good to know
Section titled “Good to know”- Per-user by design. The scope is always your own assigned steps: autopilot never reaches across users. Everyone runs their own, and coverage self-distributes across the workflow graph.
- Two structural safety bounds: it runs as you, so it can’t exceed your access, and it only ever suggests, so it can’t commit without your approval. The worst case of a misjudged step is one unwanted draft in your review queue.
- It never approves. Steps awaiting your approval are reported, never actioned: that click stays yours.
- The
INPUT NEEDEDplaceholder is uppercase on purpose. It readsINPUT NEEDED: THIS IS A PLACEHOLDER, NOT A RESULT. DO NOT APPROVE.and names the specific missing items: a request you can’t skim past, so results get read instead of rubber-stamped. It’s posted at most once per step, and it never freezes the step: the run after you supply the input, the real result is drafted. - Rubber-stamps get caught. If someone approves a placeholder without reading it, the digest calls it out under “Approved placeholders: reopen these,” and downstream steps are held so the placeholder text never propagates as real upstream content.
- Idempotent across runs. A step that already has a result or a pending draft from an earlier run isn’t re-drafted, so a daily pass won’t pile up duplicates of yesterday’s still-pending or rejected work.
- Scheduling lives outside the skill: a scheduler fires it, running as you. /coach-setup offers to register the recurring schedule (Claude Cowork, Claude Code, or ChatGPT Codex). Each run starts fresh and recovers what it needs from your suggestions and step results, not from memory.
Related
Section titled “Related”- /coach-workflow-scan: the read-only triage autopilot runs as its first phase; run it standalone for a morning brief.
- /coach-workflow-execute: the per-step executor autopilot drives for each ready step and each input request.
- /coach-setup: registers the recurring schedule that fires this skill.
- Suggestions: the human-review gate every draft lands in.