← ClaudeAtlas

frame-featurelisted

Interview the user about a feature that has no ticket behind it — read the wiki via qmd before asking anything, settle the problem and who has it and why now, check the user's wording against `wiki/domain.md`, produce a mandatory out-of-scope list and one checkable success measure, and record the decisions the framing settled (optionally as approved, intent-marked writes to `wiki/domain.md`, `wiki/architecture.md` and `wiki/adr/`) — until there is a brief whose acceptance criteria `grill-task` can drill into. Use when the user says "new feature", "start a new feature", "frame this idea", "I want to build X", "what should we build". Interactive sessions only — every step is a question a human has to answer.
bibennurbani/geekstack-flow · ★ 3 · AI & Automation · score 69
Install: claude install-skill bibennurbani/geekstack-flow
# Frame Feature The first stage of `/tcgflow-new-feature`, which dispatches `frame-feature` → `grill-task` → `plan-task`. This skill covers what `grill-task` assumes: `grill-task` drills an agreed problem down to per-subtask acceptance criteria and takes the problem, the audience and the scope boundary as already settled. Everything before that line belongs here. ## When to use this skill Invoke this skill when: - There is an idea and no ticket — *"we should let people bulk-export reports"* — and nothing downstream can act on it yet. - A conversation has produced something worth building and nobody has written down what it is, who it is for, or where it stops. - An existing task is about to be widened into what is really a new feature. **Do not use this skill** when a Jira ticket already exists — that is `/tcgflow-plan`, and the ticket is the source of truth. Not for a tool finding either: the `task-from-*` family already starts from Snyk, Cypress and Datadog, and those sources carry the framing this skill exists to elicit. To re-scope a task that is already written, use `plan-task`'s refine-existing branch, which edits it in place. ## Interactive only **A human must answer the questions.** This skill runs in an interactive session only — the same rule the browser web test follows (ADR 0041), for the same kind of reason. In a headless orchestrated run there is nobody to answer, and a brief assembled from answers the AI supplied itself is worse than no brief: it rea