idea-brief

name

idea-brief

description

Lock an ideation run into a single-source-of-truth brief (`output/IDEA_BRIEF.md`) and a replayable multi-query plan (`queries.md`). **Trigger**: idea brief, ideation brief, research ideas, brainstorm, 找 idea, 选题, 点子, 找方向. **Use when**: the user wants research ideas and their input is long / multi-turn; you need to clarify scope+constraints before retrieval. **Skip if**: the goal is to write a survey draft directly (use `arxiv-survey*` pipelines instead). **Network**: none. **Guardrail**: do not invent papers/citations; do not start retrieval here; keep the brief structured (no long prose).

Idea Brief (SSOT for ideation)

Goal: turn "a lot of user thoughts" into an auditable idea contract that downstream steps can follow without drift.

This skill does not retrieve papers. It only locks:

• scope + constraints + exclusions,
• evaluation rubric (how we will pick ideas),
• multi-query buckets (how we will search), into output/IDEA_BRIEF.md and queries.md.

Role cards (prompt-level guidance)

• Ideation PM

  • Mission: lock scope/constraints so the pipeline won't drift.
  • Do: ask only the minimum questions that materially change retrieval or feasibility.
  • Avoid: letting the run proceed with "vague topic + no constraints".

• Query Architect

  • Mission: translate scope into a multi-query plan (not one giant query).
  • Do: write 3-6 query buckets and explicit exclusions in queries.md.
  • Avoid: infinite keyword expansion; keep buckets few and purposeful.

Inputs

Required:

• GOAL.md
• DECISIONS.md
• queries.md

Optional:

• PIPELINE.lock.md (confirm the pipeline is idea-finder)
• Existing output/IDEA_BRIEF.md (refine in place if re-running)

Outputs

• output/IDEA_BRIEF.md
• queries.md
• DECISIONS.md (optional: add open questions / constraints reminders)

Procedure

1. Read GOAL.md and the current kickoff block in DECISIONS.md

• Extract the topic, the why-now, and any hard constraints/excludes already stated.

2. Ask the minimum clarification questions (max 2 rounds)

• Focus only on decisions that change the run:
  • timebox / compute / data availability
  • in-scope vs out-of-scope
  • what a "good idea" means (rubric weights)
• If the user says "你自己决定", proceed with defaults below.

3. Write or update output/IDEA_BRIEF.md (single source of truth) Must include (bullets/tables; no long prose):

• Scope (in/out)
• Constraints (timebox/compute/data/implementation appetite)
• Exclusions (>=3 or explicit "none")
• Rubric + weights (sum=1.0)
• Targets (core size, pool size, shortlist size)
• Stop rules (bounded iterations)
• Query buckets (3-6) + exclusions
• Open questions (if any)

4. Rewrite queries.md as a replayable multi-query plan Keep it compatible with retrieval scripts (literature-engineer reads - keywords: / - exclude: lists):

• Under - keywords: put bucket-level query strings (3-6 lines).
• Under - exclude: put explicit exclusions (0-10 lines).
• Set scalars:
  • - max_results: "1800"
  • - core_size: "100"
  • - time window: (optional; from/to)

Also add ideation-only config (for humans/LLM; retrieval may ignore):

• - draft_profile: "idea_finder"
• - idea_pool_min: "60"
• - idea_pool_max: "90"
• - idea_shortlist_size: "7"
• - operator_mix: "50/50" (cross-domain vs failure/eval)

5. If PIPELINE.lock.md exists, sanity-check it

• Confirm it points to pipelines/idea-finder.pipeline.md.
• If it points elsewhere, write a warning into DECISIONS.md (do not change lock silently).

Acceptance

• output/IDEA_BRIEF.md exists and is actionable (structured, not narrative).
• queries.md has non-empty - keywords: items and (optional) - exclude: items.
• queries.md sets core_size and ideation targets (pool/shortlist).
• Any unresolved ambiguity is listed under "Open questions" in output/IDEA_BRIEF.md and/or DECISIONS.md.

Troubleshooting

Issue: queries.md keywords are empty, retrieval will fail

Fix:

• Add 3-6 bucket-level query strings under - keywords: (not single tokens).
• Add exclusions under - exclude: to control noise.

Issue: brief is too vague ("topic only")

Fix:

• Add at least: constraints (timebox/compute/data) + exclusions + rubric weights.