idea-creator

What it does

Generates publishable research ideas for a given direction, then ranks and annotates them through a multi-phase process. It integrates landscape surveying, idea generation, objective feasibility checks, and a cross-model jury before selecting ideas for pilot experiments. The workflow is designed to work with additional steps like /research-lit, /novelty-check, and /research-review as part of an overall pipeline.

How it works

• Phase 0: Optional loading of a Research Wiki to establish context; otherwise, proceeds without wiki input.
• Phase 1: Landscape Survey
  • Scan local papers in papers/ and literature/ and read first 3 pages of relevant works.
  • Web search recent literature (top venues last 2 years, recent arXiv last 6 months) using 5+ query forms; read abstracts/intros of top 10-15 papers.
  • Build a landscape map, grouping by sub-directions, identifying tried approaches, limitations, and open problems.
  • Identify structural gaps such as domain transfers, contradictory findings, untested assumptions, scaling gaps, and diagnostic questions.
• Phase 1.5: Parallel lens fan-out
  • Generate ideas across multiple analytic lenses; shard output includes shard_id and candidates with dedup keys.
• Phase 2: Idea Generation
  • Use the selected reviewer backend (codex or manual) to brainstorm 8-12 concrete ideas.
  • Prompt includes the user direction, landscape map, and gaps; for each idea provide summary, core hypothesis, minimal viable experiment, contribution type, risk, and estimated effort.
• Phase 3: Mechanical consolidation + objective feasibility gate
  • Phase 3a: Drop ideas only if they exceed available GPU time (one week) or require unavailable datasets; otherwise annotate with effort_note.
  • Phase 3b: Annotate each surviving idea with prior_work after 2-3 targeted searches.
  • Phase 3c: Annotate with so_what to indicate potential impact.
• Phase 4: Deep Validation (cross-model jury)
  • Cross-model triage ranks all candidates using the chosen reviewer backend; provides objections, failure modes, novelty assessment direction, and ranking.
  • Novelty check is run on top picks only via /novelty-check workflow.
  • Select top 2-3 ideas for Phase 5 pilots.
• Phase 5: Parallel Pilot Experiments
  • Design and deploy minimal pilots (target 30 minutes to PILOT_MAX_HOURS per idea on 1 GPU).
  • Launch in parallel with run_in_background: true across GPUs; monitor with /monitor-experiment and enforce PILOT_TIMEOUT_HOURS.
  • Compare results and re-rank for next steps; track total GPU hours against MAX_TOTAL_GPU_HOURS.

When to use it

Use when the user wants to brainstorm ideas or explore publishable directions such as phrases like:

• 找idea
• brainstorm ideas
• generate research ideas
• what can we work on
• explore a research area for publishable directions

What it can touch

• Tools and prompts sections include: Bash, Read, Write, Grep, Glob, WebSearch, WebFetch, Agent, mcp__codex__codex, mcp__codex__codex-reply, mcp__manual_review__review, mcp__manual_review__review_reply
• Outputs are structured as landscape maps, candidate ideas with fields: summary, hypothesis, MVP experiments, contribution type, risk, effort, dedup_key, prior_work, so_what, effort_note

Caveats

• Phase outputs depend on external literature and reviewer backends; not all results are guaranteed reliable or novel.
• Budget and pilot constraints are enforced via constants: PILOT_MAX_HOURS, PILOT_TIMEOUT_HOURS, MAX_PILOT_IDEAS, MAX_TOTAL_GPU_HOURS, OUTPUT_DIR, and reviewer settings.
• If manual review is selected as the backend, the exact prompt text used must be preserved for follow-ups.
• The process avoids dropping ideas solely on assumed novelty or perceived impact; novelty and quality judgments occur in Phase 4 and Phase 5 through cross-model review and novelty checks.