| name | discover |
| description | Run the pre-implementation moves that reduce unknowns before planning starts - a blind spot pass over unfamiliar territory, a brainstorm/prototype pass for an undecided shape, and a one-question-at-a-time interview for ambiguities that would change architecture. Use before to-plan when the request touches unfamiliar code, an undecided UX/architecture direction, or open ambiguities; skip straight to to-plan when the work is already well understood. |
| disable-model-invocation | true |
Discover
Before a plan is worth writing, find out what you don't know yet.
This skill runs the moves that surface unknowns, ad hoc, in whatever combination the situation needs - never a fixed pipeline.
Invoking this skill is the task - start immediately, don't wait for further instructions.
Unknowns, named
Each mode targets one kind of unknown: the blind spot pass surfaces unknown unknowns (what you haven't considered), brainstorm + prototype surfaces unknown knowns (criteria you'd recognize but haven't articulated), the interview closes known unknowns (gaps you're aware of).
Known knowns - what's already in the prompt - need nothing from this skill.
Decide what this invocation needs
Read the request and judge which signals apply. Run only the modes a signal actually points to:
- Touches a codebase area, domain, or technology the user (or you) don't know well -> blind spot pass.
- Involves a UX or architecture shape that isn't decided yet, or the user asks to see options / mock something up -> brainstorm + prototype.
- Leaves ambiguities standing whose answer would change the architecture -> interview.
More than one signal can apply - run the modes in sequence (a blind spot pass often surfaces a shape decision for brainstorm/prototype, which surfaces ambiguities for the interview).
If none apply, say so plainly and recommend /to-plan directly.
Run a mode because a real unknown is sitting there, not to look thorough.
Mode 1: Blind spot pass
Explore broadly instead of narrowly answering the request as stated.
Produce: unverified assumptions the request rests on, adjacent systems or consumers that might be affected, existing prior-art patterns worth checking before inventing a new one, and open questions whose answer would change direction.
See references/modes.md for trigger signals and example prompts.
Mode 2: Brainstorm + prototype
Produce an intervention-point table (option, effort, what changes, tradeoff - ordered cheap to ambitious) so the user reacts to a spread of choices instead of the first idea.
When the work affects UI, also produce 2-4 cheap HTML mockups of wildly different directions embedded in the output artifact, for the user to react to before anything real is wired up.
Recommend the artifact-design skill for the mockups' construction quality; don't invoke it yourself.
See references/modes.md for depth.
Mode 3: Interview
Ask one question at a time with AskUserQuestion, most architecturally significant first - each answer can change what the next question should be, so batching wastes rounds on now-irrelevant questions.
Stop once remaining ambiguities are cosmetic or reversible, not architectural.
Output
- Pick a kebab-case
topic-slug naming the outcome (same convention to-plan uses for plan-name).
- Map what came out of the modes you ran onto
DISCOVERY_DATA per references/data-schema.md. Omit sections for modes you didn't run - never render an empty section.
- Write
~/tmp/{project-slug}/reports/{topic-slug}-discovery.html from templates/discovery.html, replacing only the data block between its markers.
- Publish it with the Artifact tool: same
file_path, favicon 🔍 (kept stable across redeploys of this artifact type), and a title/description built from the topic. Tell the user both the local path and the returned URL.
- Write a plain-markdown companion,
~/tmp/{project-slug}/reports/{topic-slug}-discovery-notes.md, mirroring the same content - this is what /to-plan reads back in directly, without parsing HTML.
- Recommend the user run
/to-plan next and reuse topic-slug as the plan name, so to-plan picks up {topic-slug}-discovery-notes.md automatically and treats its answers as settled rather than re-litigating them.