| name | assistant-ideate |
| description | Generate and rank options. Use light mode for quick improvement scans and deep mode for explicit brainstorming or broad exploration. |
| effort | medium |
| triggers | [{"pattern":"brainstorm|feature ideas?|what if we|how could we|possibilities for|options for|alternatives to|explore (ways|options|ideas)|what else (could|can|should)|more areas to improve|ideas for","priority":55,"min_words":4,"reminder":"This request matches assistant-ideate. Load and follow this SKILL.md and its contracts for structured brainstorming."}] |
Idea Generation
Mode-aware ideation that turns vague opportunities into proportionally ranked options.
Core principle: Match ideation depth to the decision.
Contracts
| Contract | File | Purpose |
|---|
| Input | contracts/input.yaml | Seed, mode, goal, and constraints |
| Output | contracts/output.yaml | Mode-appropriate ranking, next step or decision packet |
| Phase Gates | contracts/phase-gates.yaml | Light and deep pipeline gates |
Contract v2 makes the former deep-only ranking, refinement, and decision
artifacts conditional; deep mode preserves those v1 artifacts, while light mode
uses quick_ranking and recommended_next_step.
Rules:
- Infer light mode unless the user explicitly asks for broad/deep exploration
- Light mode produces 3-5 options and a quick ranking without a full scoring matrix
- Deep mode produces 8-15 ideas, full scoring, and at least 3 refined candidates
- Never present a single option — ideation means choices
- Constraints from input are respected throughout all phases
- Deep DECIDE output includes
decision_point and decision_options; capture user_decision only after an actual user choice
Required References
Before running the pipeline, load and apply references/ideation-pipeline.md.
Light mode uses only that compact selector. Deep mode additionally loads
references/deep-ideation.md; load it only for deep mode.
Goal
Generate several useful options at the least costly depth that supports the user's decision.
Success Criteria
- The response states whether it used light or deep mode.
- Light mode returns 3-5 options, a quick ranking, trade-offs, and one recommended next step.
- Deep mode generates at least 8 ideas before full scoring and refines at least the top 3.
- Neither mode fabricates a user choice or presents one unexamined option.
Constraints
- Do not force deep-mode ceremony onto a quick improvement or "what else" request.
- Do not collapse to one recommendation before generating multiple options.
- Ask only when missing goals or constraints would materially change the idea space or ranking.
- Respect explicit constraints throughout scoring and refinement.
- Respect prior attempts; do not re-propose rejected approaches unchanged.
- Include a wild or unconventional idea in deep mode.
Modes
Light mode
Use for quick improvement scans, "what else" questions, or a small set of
alternatives. Run UNDERSTAND → DIVERGE → CONVERGE → DECIDE with 3-5 options,
a qualitative quick ranking, trade-offs, first steps, and one recommended next
step. Do not build a five-criterion score table unless the user asks.
Deep mode
Use when the user explicitly asks to brainstorm, explore broadly, generate many
ideas, or compare a wide solution space. Run the full pipeline with 8-15 ideas,
five-criterion weighted scoring, at least 3 refined candidates, and a decision
packet.
Pipeline
Light: UNDERSTAND -> DIVERGE (3-5) -> CONVERGE (quick rank) -> DECIDE (next step)
Deep: UNDERSTAND -> DIVERGE (8-15) -> CONVERGE (weighted) -> REFINE -> DECIDE
Output
Return:
- Both modes:
ideation_mode, problem statement, and multiple ideas.
- Light:
quick_ranking and recommended_next_step.
- Deep:
ranked_ideas, refined_candidates, decision_point, and decision_options.
user_decision only after the user explicitly chooses.
Stop Rules
- Stop and ask when the goal or hard constraints are too ambiguous to generate useful options.
- In light mode, stop after the ranked options and recommended next step; do not force a follow-up decision.
- In deep mode, stop after the decision point and options unless the user already chose or asked for implementation.
- Do not enter implementation; hand selected candidates to
assistant-workflow when build work starts.
Integration with Workflow
When an idea is selected:
- Decompose into testable criteria (using
assistant-workflow's idea-to-action pipeline)
- Hand off to workflow for implementation
- Save the outcome only when the user asks or it contains a concrete durable lesson