| name | ask-human |
| layer | method |
| description | Emit a blocking question the owner can read in ~10 seconds and answer in one reply โ the scannable, recommendation-first handoff every agent posts when it hits a fork it can't own. Leads with the one decision + a recommendation, carries the ๐ค provenance header, doubles as the |
| allowed-tools | Bash, Read, Grep |
Ask the human โ the scannable, one-decision handoff
You hit a fork. First apply the decide-vs-ask test (AGENTS.md โ Deciding vs
asking): ask only when it's irreversible/expensive and not derivable and
genuinely the human's call. If any one fails โ decide + log a Decisions-log line and
keep moving. Don't reach for this skill to dodge a call you can make.
If it is ask-worthy, this skill is how you ask well. The owner may be holding ~20
async threads on their phone: the comment must be graspable in ~10 seconds,
self-contained (a fresh session resumes from it with zero memory โ #639), and it must
always propose an answer. Never post a naked "what do you want?".
The template
Post this as a top-level issue comment (add_issue_comment) โ never a PR review body
(the owner misses those, #846). One decision per comment; batch related questions into it.
## ๐ Blocked โ <the one decision, in one line>
> ๐ค **Claude Code** ยท interactive agent ยท posted from @pmcp's account (not Maarten) ยท _<one-line context>_
**TL;DR โ recommend <X>:** <the decision restated + why X is your pick, in one or two lines>
**Status:** <what's done ยท `branch-name` pushed? ยท what's NOT done>
**Why it came up:** <what cannot proceed until this is answered>
**Options:**
- **A) <label>** โ <consequence> _(recommended)_
- **B) <label>** โ <consequence>
**Reply:** `A` or `B` (or in-medium โ see below). Your reply spawns a fresh session that
resumes from THIS ticket; `lgtm`/`approve` if A-as-recommended is fine.
**Don't lose:** <decisions/assumptions the next agent must carry forward>
Rules that make it work:
- Lead with the decision + a recommendation. The first two lines must answer "what's
being asked, and what does the agent suggest?" with no scrollback.
- Always a recommendation. You did the work; you have an opinion โ state it and why.
- Provenance header is mandatory (enforced by
require-comment-provenance). Interactive
agents post under @pmcp's account โ use the "not Maarten" disclaimer above. A bot-account
pipeline comment uses > ๐ค **<tool>** ยท agent pipeline (CI) ยท _<context>_ instead (no
@pmcp disclaimer โ it'd be false).
- @mention only because action is needed. This is an ask โ
@mention @pmcp
(NOTIFY_HANDLE). Pure FYIs get no mention.
- Push before you block. If you've written anything,
git push -u origin <branch> first
and name that branch under Status โ an unpushed worktree is lost on stop (#639).
- Then apply
status:blocked and stop.
Attach the right medium
Prose is the fallback, not the default. If the question is visual or structural, a
paragraph is the wrong channel โ the owner shouldn't have to reconstruct a layout or a data
model in their head. Attach the artifact that shows it, and let them reply on the artifact
itself (the reply loop is WS5, #1191). We already own every medium โ pick by what the question
is, not by habit.
| Your question is aboutโฆ | Attach (medium) | Produce it with | Owner replies by |
|---|
| How it looks / feels (a rendered UI, spacing, copy, states) | live preview | ui-proposal (staging deploy, NUXT_PUBLIC_CROUTON_REVIEW=true) โ or a screenshot if no runnable app | pinning feedback on the preview โ ๐ฏ Preview feedback comment naming the file; or A/B |
| One screen / one state (a single view, an error, a before/after) | screenshot | node scripts/app-shots.mjs <baseUrl> <path[:name]> โ screenshots/<name>.png | commenting / A/B |
| A flow or interaction (multi-step, timing, motion) | short video | demo-video (WebM storyboard) | commenting |
| The data model (fields, types, relationships) | schema render | schema-review (โ PNG/HTML/MD) | inline comment on the committed .md; or A/B |
| Structure / status / dependencies (what depends on what, where the tree is) | diagram | ticket-diagram (Excalidraw on the epic) | editing the Excalidraw โ scripts/ticket-excalidraw-import.mjs round-trips it back; or A/B |
| A tradeoff / priority / naming (no visual or structural surface) | prose | the block above | A/B/lgtm |
Rules:
- One artifact, the most direct one. Don't attach three renders "to be safe" โ pick the
medium that answers this question fastest and link it from the blocking comment.
- The artifact supplements the block; it doesn't replace it. Still lead with the decision +
recommendation in text โ the artifact is the evidence, the block is the ask.
- Screenshots land in
screenshots/ (gitignored) โ never the repo root or an app dir.
- Reuse the skill, don't reinvent it. Each medium is its own skill/script with its own
gotchas; invoke it, don't hand-roll a render.
- If producing the artifact would cost more than the answer is worth (a deploy for a one-line
copy tweak), fall back to a screenshot or prose โ match the effort to the stakes.
Why this shape
- 10-second-scannable โ the owner triages 20 threads without opening each.
- Recommendation-first โ most replies collapse to
lgtm, one round-trip (epic metric #1).
- Self-contained โ the resuming session (
resume-on-comment.yml, fresh, checks out
main) continues from the branch + comment without re-deriving or diverging (#639).
This extends #639 (the handoff block) โ it doesn't replace it. #639 gave the block its
state/after/don't-lose fields; this adds the scannable lead + the always-a-recommendation
rule and packages it so every agent emits the same shape.
