Collaborative Brainstorming

Turn uncertainty into a useful choice or a focused experiment.

The user's instructions take precedence over this skill; carry existing authorization forward and do not add approval gates.

Choose the Work Shape

Ground the Question

Recall prior decisions through the installed Sibyl skill when configured (sibyl context "<decision>" --intent plan), and recheck any volatile claim it returns. Find the nearest existing instance and the prior work before generating alternatives.

Translate taste into examples and observations. For "smooth," identify the interaction and visible delay. For "gorgeous," inspect the reference and name the composition, motion, or typography worth preserving. Quantification helps where a metric represents the experience; a proxy metric does not replace the experience itself. A reference the user supplies is the specification: match it before reinterpreting it, and when a concept could move, show it moving.

Treat the user's hunch as a hypothesis: a decisive check can confirm part of it and reject the rest, so report both. Do not downgrade an explicit "use X" to a preference because it was written casually; if it conflicts with feasibility, say so before designing around it.

Elicit what the user knows but did not write. A prompt is a map of the problem, and the user holds constraints, taste, and context they treat as obvious; a wrong guess at those costs a rebuilt slice. When the user signals ambiguity or a picture they cannot describe, interview them: a chain of specific, answerable questions through the host's question tool, or a rendered artifact that puts each question beside the diagram, snippet, or option it concerns. Stop when answers stop changing the design. For taste-driven work, distinct rendered directions are the interview: name the style of each so the user gains the vocabulary to steer, and treat the first pass as a probe. When the user says they do not know the domain, open with the questions they did not ask, ranked by consequence, before any recommendation.

Compare the Alternatives That Matter

Add an unconventional option only when the obvious ones share an untested assumption.

Test the Uncertainty Before Polishing

Pick the next move by its power to change the decision; a rough prototype often settles more than another design paragraph.

Reviewer agreement is not proof, since reviewers may share training, sources, or framing. Settle consequential disagreements with code, sources, or an experiment, and say when evidence stays inconclusive.

Decide and Preserve the Reason

Recommend the best-supported direction with its main tradeoff and next action, or keep exploration open when that is the deliverable.

For a durable decision, record the choice, rationale, rejected alternative worth remembering, evidence, and reconsideration trigger in configured project memory. Preserve deliberately open decisions and user corrections verbatim when paraphrase would narrow them.

Route unresolved factual questions to research, complex execution to plan, and a clear bounded change to implement; these are optional compositions, not a mandatory pipeline.

Evidence

Reviewed 2026-09-04 and 2026-10-27; sources and their limits are in references/evidence.md.

Anti-Patterns

What This Skill is NOT

• A prerequisite for implementation, debugging, or routine fixes.
• A requirement to produce a spec, a fixed option count, or a final approval question.
• A substitute for the user's product judgment or current technical evidence.