brycewang-stanford/scientify-idea-generation

Idea Generation

Don't ask permission. Just do it.

Generate innovative research ideas grounded in literature analysis. This skill reads existing papers, identifies research gaps, and produces 5 distinct ideas with citations.

Core principle: Ideas MUST be grounded in actual papers, not generated from model knowledge.

Workspace: See ../_shared/workspace-spec.md for directory structure. Outputs go to $WORKSPACE/ideas/.

Step 1: Check Workspace Resources

First, check what resources already exist:

# Check active project
cat ~/.openclaw/workspace/projects/.active 2>/dev/null

# Check papers
ls ~/.openclaw/workspace/projects/*/papers/ 2>/dev/null | head -20

# Check survey results
cat ~/.openclaw/workspace/projects/*/survey/clusters.json 2>/dev/null | head -5

Assess Available Resources

| Resource | Location | Status | |----------|----------|--------| | Papers | $WORKSPACE/papers/ | Count: ? | | Survey clusters | $WORKSPACE/survey/clusters.json | Exists: Y/N | | Repos | $WORKSPACE/repos/ | Count: ? |

Step 2: Ask User About Search Strategy

Based on workspace state, ask user:

If papers exist (>=5):

Found {N} papers in workspace from previous survey.

Options:

1. Use existing papers - Generate ideas from current collection
2. Search more - Run /literature-survey to expand collection
3. Quick search - Add 5-10 more papers on specific topic

If no papers:

No papers found in workspace.

To generate grounded ideas, I need literature. Options:

1. Run /literature-survey - Comprehensive search (100+ papers, recommended)
2. Quick search - Fetch 10-15 papers on your topic now
3. You provide papers - Point me to existing PDFs/tex files

Step 3: Acquire Resources (if needed)

Option A: Delegate to /literature-survey (Recommended)

If user wants comprehensive search:

Please run: /literature-survey {topic}

This will:
- Search 100+ papers systematically
- Filter by relevance (score >=4)
- Cluster into research directions
- Save to $WORKSPACE/papers/

After survey completes, run /idea-generation again.

Option B: Quick Search (5-10 papers)

For fast iteration, do minimal search:

1. ArXiv search:

Tool: arxiv_search
Arguments:
  query: "{user_topic}"
  max_results: 10

2. Clone 3-5 reference repos:

mkdir -p $WORKSPACE/repos
git clone --depth 1 {repo_url} $WORKSPACE/repos/{name}

3. Download paper sources:

mkdir -p $WORKSPACE/papers/{arxiv_id}
curl -L "https://arxiv.org/src/{arxiv_id}" | tar -xz -C $WORKSPACE/papers/{arxiv_id}

Step 4: Analyze Literature

Prerequisites: At least 5 papers in $WORKSPACE/papers/

4.1 Read Papers

For each paper, extract:

• Core contribution (1 sentence)
• Key method/formula
• Limitations mentioned
• Future work suggestions

Long papers (>50KB): See references/reading-long-papers.md

4.2 Identify Research Gaps

Look for:

• Common limitations across papers
• Unexplored technique combinations
• Scalability issues
• Assumptions that could be relaxed

Document gaps in $WORKSPACE/ideas/gaps.md:

# Research Gaps Identified

## Gap 1: [Description]
- Mentioned in: [paper1], [paper2]
- Why important: ...

## Gap 2: [Description]
...

Step 5: Generate 5 Ideas

Create $WORKSPACE/ideas/idea_1.md through idea_5.md using template in references/idea-template.md.

Requirements:

• Each idea cites >=2 papers by arXiv ID
• Use different strategies:

| Idea | Strategy | |------|----------| | 1 | Combination - merge 2+ techniques | | 2 | Simplification - reduce complexity | | 3 | Generalization - extend to new domain | | 4 | Constraint relaxation - remove assumption | | 5 | Architecture innovation - new design |

REJECTED if: No arXiv IDs cited, or ideas not grounded in literature

Step 6: Select and Enhance Best Idea

6.1 Score All Ideas

| Idea | Novelty | Feasibility | Impact | Total | |------|---------|-------------|--------|-------| | 1 | /5 | /5 | /5 | /15 | | ... | | | | |

6.2 Enhance Selected Idea

Create $WORKSPACE/ideas/selected_idea.md with:

• Detailed math (loss functions, gradients)
• Architecture choices
• Hyperparameters
• Implementation roadmap

6.3 (Optional but recommended) OpenReview Evidence Check

For the top 1-2 shortlisted ideas, validate novelty/positioning risk with openreview_lookup:

• Query using core title keywords or representative baseline paper title
• Extract evidence:
  • decision (if available)
  • average rating/confidence
  • reviewer weakness patterns
• Add a short "submission risk note" section per idea:
  • likely reviewer concern
  • mitigation experiment to add
  • positioning adjustment

Do not claim accept/reject predictions as facts. Report evidence-backed risk signals only.

Step 7: Code Survey

Map idea concepts to reference implementations.

See references/code-mapping.md for template.

Output: $WORKSPACE/ideas/implementation_report.md

Step 8: Summary

Create $WORKSPACE/ideas/summary.md:

• All 5 ideas with scores
• Selected idea details
• Next steps: /research-pipeline to implement

Commands

| User Says | Action | |-----------|--------| | "Generate ideas for X" | Check workspace -> ask strategy -> generate | | "I have papers, generate ideas" | Skip to Step 4 | | "Enhance idea N" | Jump to Step 6 | | "Map to code" | Jump to Step 7 |

Integration

• Before: /literature-survey to collect papers
• After: /research-pipeline to implement selected idea
• Alternative: /write-review-paper to write survey instead