| name | idea-pool-expander |
| description | Generate a large, structured brainstorm pool (60-90) for ideation and write it into `output/IDEA_SHORTLIST.md`.
**Trigger**: idea pool, brainstorm pool, expand ideas, 发散, 头脑风暴, idea candidates.
**Use when**: you have a core set + taxonomy map and want to expand thinking before curating a shortlist.
**Skip if**: you only want a small, feasibility-first shortlist (use `idea-shortlist-curator` directly).
**Network**: none.
**Guardrail**: no invented papers; each Tier-1/2 idea must point to `papers/core_set.csv` paper_ids; keep cards short.
|
Idea Pool Expander (operator-driven)
Goal: create a big pool of idea candidates that genuinely expands thinking, but remains controllable and auditable.
This skill writes the Idea Pool section of output/IDEA_SHORTLIST.md.
The final shortlist is handled by idea-shortlist-curator.
Inputs
output/IDEA_BRIEF.md
outline/taxonomy.yml
papers/paper_notes.jsonl
papers/core_set.csv
Outputs
Operators (use explicitly)
Use these operator families to force diversity. Target >=6/8 families represented.
- Counterfactual / constraint flip
- Failure-mode-first
- Evaluation protocol swap (task/metric/budget)
- Component swap in an agent loop
- Combination with explicit assumptions
- Cross-domain analogy import
- Negative-result mining
- System/product constraints (cost/latency/auditability)
Procedure (best-of-N, then write)
- Read the contract from
output/IDEA_BRIEF.md
- Scope / exclusions / constraints
- Targets: pool size range, shortlist size, operator mix (default 50/50)
- Use
outline/taxonomy.yml as the idea map
- Ensure the pool covers multiple clusters (avoid all ideas collapsing into one bucket).
- Use
papers/paper_notes.jsonl to avoid fantasy ideas
- Prefer ideas that respond to explicit limitations/failure modes recorded in notes.
- Multi-sample within each operator family (best-of-N)
- For each operator family, generate 2-3 alternative mini-sets.
- Choose the best mini-set by: novelty, clarity, falsifiability.
- Write the Idea Pool into
output/IDEA_SHORTLIST.md
- Pool size: 60-90 (hard min 60)
- Each idea is a short card (3-6 lines):
- Tier label (Tier-0 Wild / Tier-1 Plausible / Tier-2 Ready)
- Operator tag
- One-liner problem statement
- Key assumption
- How to falsify (1 sentence)
- Evidence pointers:
- Tier-0: optional
- Tier-1/2: required (>=1
paper_id from papers/core_set.csv)
- Add a small "Pool Diagnostics" block
- Counts by tier
- Counts by operator family
- Coverage across taxonomy clusters
Acceptance
output/IDEA_SHORTLIST.md contains an "Idea Pool" section with 60-90 ideas.
- Operator coverage >=6/8.
- Every Tier-1/2 idea includes at least one pointer to
papers/core_set.csv (paper_id).
- Cards are short; no long paragraphs.
Troubleshooting
Issue: pool feels repetitive / not expanding
Fix:
- Increase operator coverage (force missing families).
- Add at least 5 ideas that are failure-mode-first and 5 that are evaluation-protocol-swap.
Issue: too many wild ideas
Fix:
- Convert some Tier-0 into Tier-1 by attaching at least one
paper_id pointer.
- Or explicitly tag as "needs verification" and keep it out of the final shortlist.