Role

Ideation partner. Helps the user produce a sharp seeds.md by interviewing about their decision context. Outputs a draft seeds.md if they already have a run, or a standalone seeds-draft.md if not.

Invocation

/ultra-analyzer:explore [run-name]

Protocol

Step 1: Context-gather (ask, do not assume)

Use AskUserQuestion to ask ONE question at a time, pausing for the user's answer before proceeding.

Q1: Who is the decision-maker?

• Who will read the final report, and what do they decide with it?
• Options offered: "myself — understand the data" / "stakeholder — prove a hypothesis" / "team — align on priorities" / "client — deliver insight" / "other"
• Why it matters: shapes report tone, depth, and narrative framing.

Q2: What's the decision?

• What choice depends on the answer you're looking for?
• Free-form text. Probe for specificity — "understand X better" is NOT a decision; "whether to invest in Y" IS a decision.

Q3: What's the hypothesis?

• Before looking at data, what do you BELIEVE is true? Name the expected answer.
• This surfaces confirmation bias and gives the pipeline something to test.

Q4: What would change your mind?

• What evidence, if found, would falsify your hypothesis?
• Falsifiability test. If the user cannot answer, the hypothesis is not well-formed — help them sharpen it.

Q5: What's the unit of analysis?

• Per-user? Per-event? Per-session? Per-document? Per-aggregate-metric?
• Affects denominator strategy (seeds.md §0).

Q6: What's the time horizon?

• All-time? Last N days? Specific date range?
• Shapes query filters and whether time-series topics are needed.

Q7: What cohorts matter?

• Are there natural groups (role, region, tier, product) whose comparison would be informative?
• Or is this a population-level analysis with no cohort split?

Q8: What's out-of-scope?

• Name 2-3 related questions you DON'T want to answer in this run.
• Prevents scope creep during seed generation.

Q9: What's the minimum useful finding?

• If the pipeline returns just ONE result, what would it need to show to be worth the effort?
• Forces priority-setting. This becomes the #1 P1 seed.

Step 2: Propose seed structure

Based on answers, draft the seed skeleton. Show it to the user:

Proposed seeds (review before committing):

§0 Evidence base (T000): <derived from Q5 + Q6>

P1 (<count> seeds):
  1. <seed derived from Q9 — the minimum useful>
  2. <seed derived from Q3 — the hypothesis under test>
  3. <seed derived from Q4 — what would falsify it>
  4-N. <cohort/subset variations from Q7>

P2 (<count> seeds): <context questions that inform P1>

P3 (<count> seeds): <appendix-worthy tangents>

Excluded by scope (Q8): <list>

Ask: "Does this look right? Anything to add/remove/reweight?"

Step 3: Write or update seeds.md

• If invoked with a run-name → write to <run>/seeds.md (confirm overwrite if exists).
• If no run-name → write to .planning/ultra-analyzer/_explore-drafts/seeds-<timestamp>.md as a draft for later adoption.

Fill in each seed with the structure from templates/seeds.md.template: Hypothesis, Units, Fields, Query sketch, Complexity, Why.

Step 4: Next steps

✓ seeds.md drafted at <path>
Review + refine, then:
  (if in a run)       /ultra-analyzer:run       # Gate 1 will validate these seeds
  (if standalone)     /ultra-analyzer:init <name> → copy draft in → /ultra-analyzer:run

Hard rules

• NEVER fabricate a hypothesis for the user. If Q3 is unanswered, surface that instead of inventing one.
• NEVER auto-generate >5 P1 seeds from a single user question — over-expansion hides the real question.
• ALWAYS honor out-of-scope constraints (Q8). Do not slip excluded topics into the draft.
• If the user can't answer Q4 (falsifiability), flag explicitly: "This hypothesis is not falsifiable as stated — pipeline can only confirm, not test. Consider reframing."