Skip to content

idea-brainstorm.md

Latest commit

 

History

History
186 lines (126 loc) · 6.78 KB

idea-brainstorm.md

File metadata and controls

186 lines (126 loc) · 6.78 KB
Raw

Idea Brainstorm Guide

Languages: English | 简体中文

Navigation: Project README | 项目主页

1. What This Workflow Is For

idea-brainstorm is for turning a topic into a literature-grounded research direction memo.

It is designed for situations like:

  • “help me find promising thesis directions”
  • “what should we discuss with the advisor next”
  • “map the landscape before deciding what to build”

It is not a survey-writing pipeline and not a project-spec generator. The goal is a compact, discussion-ready memo, not a full paper draft and not an execution plan.

2. Terminal Deliverable

The default reader-facing output is:

  • output/REPORT.md

The pipeline also expects:

  • output/APPENDIX.md
  • output/REPORT.json
  • output/IDEA_SCORECARD.md
  • output/IDEA_SCORECARD.json

Trace artifacts stay under output/trace/, so the final memo can stay relatively clean while the intermediate reasoning remains auditable.

3. What Makes It Different From Survey Writing

The idea pipeline shares the same artifact-first philosophy as the survey path, but it optimizes for a different endpoint.

Workflow Main question Main output
arxiv-survey / arxiv-survey-latex how should a literature area be synthesized into a paper draft and optional PDF
idea-brainstorm what literature-grounded research directions are worth discussing next memo plus shortlist

The difference matters:

  • the idea workflow keeps the middle stages table-first
  • it emphasizes tensions, missing pieces, and promising axes
  • it intentionally stops before turning directions into project plans or proposal prose

4. Default Shape Of A Run

The current pipeline defaults are:

  • core_size=36
  • max_results=240
  • evidence_mode=abstract
  • direction pool size 12-24
  • shortlist size 3-5
  • final report lead directions 3

That means the pipeline keeps enough literature to expose competing signals without paying survey-scale context costs before focus is locked.

5. Stage Flow

Stage Purpose Main outputs
C0 define the idea brief and get human approval output/trace/IDEA_BRIEF.md, queries.md, DECISIONS.md
C1 retrieve literature and build the core set papers/papers_raw.jsonl, papers/core_set.csv, papers/retrieval_report.md
C2 build a literature landscape and choose focus lenses outline/taxonomy.yml, updated DECISIONS.md
C3 convert papers into signal tables papers/paper_notes.jsonl, output/trace/IDEA_SIGNAL_TABLE.md
C4 generate and screen candidate directions output/trace/IDEA_DIRECTION_POOL.md, output/trace/IDEA_SCREENING_TABLE.md
C5 converge, write, score, and emit a bounded repair plan output/trace/IDEA_SHORTLIST.md, output/REPORT.md, output/APPENDIX.md, output/REPORT.json, output/IDEA_SCORECARD.json

6. The Key Artifacts

These are the files that actually matter when you inspect or debug a run:

Artifact Why it matters
output/trace/IDEA_BRIEF.md single source of truth for topic, constraints, exclusions, audience, and query framing
papers/core_set.csv defines the literature base the later brainstorm is allowed to lean on
outline/taxonomy.yml represents the idea landscape, not a paper outline
output/trace/IDEA_SIGNAL_TABLE.md where tensions, missing pieces, and promising axes become explicit
output/trace/IDEA_DIRECTION_POOL.md the first broad set of possible directions
output/trace/IDEA_SCREENING_TABLE.md makes the shortlist decision auditable instead of purely stylistic
output/trace/IDEA_SHORTLIST.md the final direction shortlist before memo writing
output/REPORT.md the final discussion memo
output/IDEA_SCORECARD.json machine-readable verdict for trace consistency, actionability, diversity, and literature anchors

The scorecard validates the observable memo contract. It does not claim that a direction is scientifically novel; that still requires deeper reading and expert judgment.

7. How To Run It

Typical prompt:

Brainstorm literature-grounded research ideas around embodied agents for home robotics.

If you want to pin the pipeline:

Use idea-brainstorm to generate a research direction memo on embodied AI.

If you want the pipeline to respect a specific audience:

Use idea-brainstorm and optimize the memo for advisor discussion, not project planning.

8. Core Skills Behind The Workflow

The main behavior comes from a small chain of idea-specific skills:

  • idea-brief
  • literature-engineer
  • dedupe-rank
  • taxonomy-builder
  • idea-signal-mapper
  • idea-direction-generator
  • idea-screener
  • idea-shortlist-curator
  • idea-memo-writer
  • deliverable-selfloop
  • artifact-contract-auditor

The important design choice is that direction generation happens after there is already:

  • a scoped brief
  • a literature base
  • a landscape taxonomy
  • a signal table

That is what keeps the memo from collapsing into vague idea dumping.

9. What Good Output Looks Like

A good brainstorm memo should be:

  • grounded in real literature rather than generic opportunity claims
  • small enough to discuss in one meeting
  • diverse enough that the shortlist is not three versions of the same bet
  • honest about uncertainty
  • clear about what each direction is trying to fix or open up

The direction pool should feel like a set of distinct academic directions, not a combinatorial list of minor prompt variations.

10. Common Failure Modes

10.1 The directions all sound the same

Usually the signal table is too shallow or the screening stage is not enforcing enough diversity. Inspect:

  • output/trace/IDEA_SIGNAL_TABLE.md
  • output/trace/IDEA_DIRECTION_POOL.md
  • output/trace/IDEA_SCREENING_TABLE.md

10.2 The memo reads like a proposal before evidence exists

That usually means the direction generation stage is over-expanding. The fix is not more prose. The fix is to keep C3 and C4 tighter and more literature-grounded.

10.3 The shortlist is flashy but not discussable

Check whether the screener over-weighted novelty language and under-weighted discussion value, evidence grounding, or thesis potential.

10.4 The output should have been a survey instead

If the real goal is to synthesize the field rather than pick directions, use the survey workflow instead:

11. When To Use Another Workflow

Do not use idea-brainstorm when:

  • you already know you need a literature review paper
  • the deliverable must be a PDF manuscript
  • you are reconstructing an existing thesis project

In those cases: