A terse design prompt ("the dropdown hover doesn't work") is usually answerable but not complete — the missing pieces surface mid-task and cost time. This runs a ~60-second completeness check first: resolve what the repo already tells you, surface the few genuine gaps, confirm which assets (if any) are actually needed — so the build/fix is fast and right the first time.
Interview only for gaps that change what you do or how fast. If you can already state the scope, the expected vs. actual (bug) or the target state + acceptance (change), and you know where the spec/assets are, skip the questions, restate the task in one line, and proceed. Interrogating for things the repo answers — or for low-stakes details — is the anti-goal. Most prompts pass the gate; the interview is the exception, not the toll booth.
When the prompt was pasted as a piece of work, the tier comes before this whole skill runs. The first reply is a Tier: line (per ~/.claude/CLAUDE.md) naming the model, reasoning label and mechanism for the build or review being asked for, plus at most one line saying what job you read in the paste, and nothing else — no repo reading, no intake questions, no code. Then wait for the maintainer's go (g) and run this skill from step 1. The intake questions come after the go, not beside the tier, because step 1's whole point is self-answering from the repo first and that reading is the spend the wait protects.
One gate, once: for design tasks this skill is the intake step — don't stack another start-of-task gate on top of it (the general form of that rule now lives in ~/.claude/CLAUDE.md, "Reach for a skill before improvising"). Run this, then hand off.
docs/.git log / blame the component or area; a recent change often is the cause.navigation/ia.ts, navigation/webRoutes.ts, the screen file, theme/primitives.tsx.)theme/tokens.ts), a theme helper, or theme/primitives.tsx changes every surface that consumes it — its blast radius is app-wide even when the task named one screen. If shared: name the other affected surfaces up front, add at least one to the manual-QA set, and plan a PR tradeoff callout. (Flagship: the #194 contrast fix darkened shared faint/muted tokens site-wide, not just the signed-out home.)just up) or read the code. For a bug, a repro beats a mockup.Bug ("X doesn't work"):
Build / change (from a mockup):
docs/design/ui-copy-guide.md, .claude/rules/grounded-answers.md.)Put the genuine unknowns to the user in one AskUserQuestion call (≤ 4 questions, each with a recommended default so it's one click). Never ask what step 1 already answered. If nothing material is missing, say "prompt is complete — proceeding" and move on.
State the sharpened task in one line — scope · deliverable · acceptance · assets needed — then continue with design-build (build/change) or systematic-debugging (bug).
Interrogating for repo-discoverable facts · asking low-value questions · gating an already-complete prompt behind an interview · requesting the whole design bundle when one state's screenshot/spec section suffices · skipping straight to code on a genuinely ambiguous prompt.
.claude/skills/design-build/SKILL.md (the build workflow this feeds) · .claude/skills/design-review/SKILL.md (the pre-build mockup evaluation that precedes intake) · .claude/rules/workflow.md rule 14 (surface real ambiguity; don't gate routine work) · docs/product-onboarding/site-navigation-guide.md. Origin: retro on the nav-dropdown hover fix (#171).
alethical-org/alethical
a84ed3e
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.