Idea-to-Proposal Skill

Overview

This skill takes a raw research idea through a rigorous multi-phase pipeline:

1. Idea Structuring — Parse any input format into a structured research brief
2. Literature Search + 8-Dimension Evaluation — Parallel agents search literature and score the idea
3. Autonomous Refinement Loop — Iteratively improve weak dimensions until all meet threshold
4. Human Review Gate — Present scores and refined idea for user approval
5. Proposal Generation — Parallel agents write method design and experiment plan

The output is a publication-quality research proposal suitable for a top-tier AI conference, saved to the user's Obsidian vault.

Setup

Before running the skill, install dependencies:

pip install -r "BASE_DIR/scripts/requirements.txt"

Required packages:

• arxiv — arXiv API access
• requests — HTTP requests for paper fetching
• bibtexparser — BibTeX parsing and generation
• semanticscholar — Semantic Scholar API integration

Parameters

Parse the following from $ARGUMENTS:

Parameter	Default	Description
--target-venue	none (general)	Conference/journal: NeurIPS, ICML, CVPR, ACL, AAAI, ICCV, ICLR
--threshold	7	Minimum score (1-10) across all 8 dimensions to pass
--max-rounds	3	Maximum autonomous refinement rounds before human gate
--domain	auto-detect	Research domain for literature search focus

Everything else in $ARGUMENTS is the idea input (text, file path, or URL).

Output Directory

~/Library/Mobile Documents/iCloud~md~obsidian/Documents/My_note/10_Projects/Idea-Proposals/YYYY-MM-DD-<topic>/
├── idea-evolution.md          # Version history through refinement rounds
├── evaluation-report.md       # Final scores + per-dimension analysis
├── proposal.md                # The full research proposal
├── references.bib             # All collected BibTeX citations
├── experiment-benchmarks.md   # Detailed benchmark/baseline survey
└── figures/                   # Mermaid diagrams

Collision handling: If directory exists, create versioned copy (-v2, -v3, etc.).

Output Quality Targets

• Proposal length: 10-15 pages (Markdown), approximately 5000-8000 words
• Target format: Conference pre-proposal or detailed concept paper
• Assumed reader: Program committee member from neighboring subfield
• Math depth: Key equations with intuition; full proofs only if central to contribution
• Citation style: [Author et al., Year] with corresponding references.bib

---

Phase 1: Idea Structuring

Purpose: Parse any input format into a structured research brief.

Input Detection

Detect input type from $ARGUMENTS:

• File path — Ends in .md, .txt, or is an absolute/relative path to a file
  • Read file content, extract the idea description
• arXiv URL — Contains arxiv.org/abs/
  • Extract paper ID, fetch metadata via paper_search.py --arxiv-id <ID>
  • Treat as paper-extension input: user wants to extend/improve this paper
• PDF path — Ends in .pdf
  • Read with pagination (pages 1-20); extract core concepts
  • Treat as paper-extension input
• Free text — Any other format
  • Parse directly as idea description

Structured Brief Schema

Normalize all inputs into this schema:

## Research Idea Brief
- **Problem Statement**: What gap/limitation exists?
- **Proposed Approach**: Core technical idea (1-3 sentences)
- **Key Innovation**: What's novel vs. prior work?
- **Target Domain**: e.g., multi-agent systems, NLP, CV, RL
- **Target Venue**: (from --target-venue or "General")
- **Input Type**: free-text | structured | paper-extension
- **Reference Papers**: (if provided, extracted metadata)

Completeness Rules

• Required: Problem Statement, Proposed Approach, Target Domain
• Optional: Target Venue, Reference Papers, Key Innovation
• If Key Innovation missing: infer from Proposed Approach
• If Target Venue missing: prompt user; if skipped, set to "General (all venues)"
• If Reference Papers missing: leave blank; populated from Phase 2

Workflow

1. Detect input type
2. If paper-extension: read paper, extract core concepts (following paper-triggered-survey patterns)
3. Normalize into structured brief schema
4. If any required field is unclear, ask user for clarification
5. Present brief to user for confirmation
6. Create output directory; save brief as Version 0 in idea-evolution.md

Wait for user confirmation before proceeding to Phase 2.

---

Phase 2: Literature Search + Multi-Dimensional Evaluation

Launch two parallel agents via the Agent tool.

Agent A: Literature Search

Agent prompt must include:

• The structured brief from Phase 1
• Instructions to use paper_search.py

Agent A workflow:

1. Generate 5-8 search queries from the brief:
   • Method-focused queries (technique + domain)
   • Problem-focused queries (problem statement keywords)
   • Cross-domain queries (technique applied in other domains)
2. For each query, run:

   python "BASE_DIR/scripts/paper_search.py" \
     --query "<query>" \
     --max-results 20 \
     --output json \
     --sort relevance
   

3. Deduplicate results across queries
4. Identify seminal works (5-year window) vs. recent advances (1-2 years)
5. Classify papers into 3-6 thematic clusters
6. Extract SOTA baselines and benchmark datasets for the target task
7. For each key paper, fetch BibTeX:

   python "BASE_DIR/scripts/bibtex_utils.py" \
     fetch \
     --title "<paper-title>"
   

8. Return structured results: paper list with themes, benchmarks, baselines, gaps

Graceful degradation:

• 0-2 papers found: Flag as "Highly Novel / Unexplored Area"
• 3-10 papers: Proceed with reduced context
• For niche ideas: identify related task benchmarks from adjacent domains

Agent B: 8-Dimension Evaluation

Agent B runs AFTER Agent A completes (needs literature context for accurate Novelty and Positioning scores).

Score the idea 1-10 on each dimension using the evaluation rubric (see references/evaluation-rubric.md):

Dimension	Key Question
Novelty	Does prior work already solve this? How differentiated from closest work?
Technical Soundness	Is the method theoretically grounded? Any logical gaps?
Feasibility	Data/compute/time realistic? For agent-based proposals: estimate token cost (model, calls/run, input/output tokens, $/run)
Significance	How important is the problem? Size of potential impact?
Clarity	Is the idea unambiguous? Could someone implement from this description?
Experimental Validity	Can claims be tested? Are metrics well-defined?
Scalability	Does approach generalize beyond the specific setup?
Positioning	Fits target venue? Right timing for the community?

Output format:

## Evaluation Round 1
| Dimension | Score | Justification |
|-----------|-------|---------------|
| Novelty | X/10 | [2-3 sentence justification citing specific papers] |
| Technical Soundness | X/10 | [justification] |
| Feasibility | X/10 | [justification; token cost if agent-based] |
| Significance | X/10 | [justification] |
| Clarity | X/10 | [justification] |
| Experimental Validity | X/10 | [justification] |
| Scalability | X/10 | [justification] |
| Positioning | X/10 | [justification] |

**Overall: X.X/10**
**Weakest dimensions:** [list dimensions below threshold]
**Refinement targets:**
- [Dimension]: [specific suggestion for improvement]
- [Dimension]: [specific suggestion for improvement]

Save evaluation to evaluation-report.md.

---

Phase 3: Autonomous Refinement Loop

Goal: Improve the idea until ALL dimensions score ≥ threshold (default 7/10).

Loop Logic

for round in 1..max_rounds:
    1. Identify dimensions scoring < threshold
    2. For each weak dimension, generate specific refinement actions:
       - Novelty low       → Differentiate from closest prior work, add unique component
       - Tech Soundness low → Add theoretical justification, fix logical gaps
       - Feasibility low    → Simplify architecture, reduce data/compute needs
       - Significance low   → Strengthen problem motivation, broaden impact scope
       - Clarity low        → Rewrite ambiguous parts, add concrete examples
       - Exp Validity low   → Add metrics, design stronger evaluation plan
       - Scalability low    → Generalize approach, remove domain-specific assumptions
       - Positioning low    → Adjust framing for target venue, cite venue-relevant papers
    3. Apply refinements to structured brief → produce updated version
    4. Re-evaluate all 8 dimensions on updated version
    5. Append to idea-evolution.md:
       - Round number
       - Before/after scores for each dimension
       - Changes made with reasoning
    6. If ALL dimensions ≥ threshold → break loop

Refinement Rules

• Conflict resolution: If refinements for different dimensions conflict, prioritize the dimension with the largest gap from threshold
• Additive only: Never contradict a refinement the user previously approved
• Structural limitations: If a dimension is stubbornly low after 2 rounds, flag it as a structural limitation rather than over-engineering a fix

Literature Context Update Policy

• If refinement modifies < 30% of technical approach: reuse Phase 2 literature context
• If refinement substantially pivots (≥ 30% change): trigger 1-2 supplementary search queries on new components only
• Log which policy was applied each round

---

Phase 4: Human Review Gate

Present the following to the user and wait for their response:

## Idea Refinement Complete — Review Required

### Final Scores (Round N)
| Dimension | Score | Status |
|-----------|-------|--------|
| Novelty | X/10 | ✓/✗ |
| Technical Soundness | X/10 | ✓/✗ |
| Feasibility | X/10 | ✓/✗ |
| Significance | X/10 | ✓/✗ |
| Clarity | X/10 | ✓/✗ |
| Experimental Validity | X/10 | ✓/✗ |
| Scalability | X/10 | ✓/✗ |
| Positioning | X/10 | ✓/✗ |
| **Overall** | **X.X/10** | **Ready / Not Ready** |

### Evolution Summary
- Round 0 → 1: [changes and score deltas]
- Round 1 → 2: [changes and score deltas]
- Round 2 → 3: [changes and score deltas]

### Structural Limitations (if any)
- [Dimension]: Capped at X/10 because [reason]

### Current Idea (Refined Version)
[Full structured brief — latest version]

### Options
- **proceed** — Generate full proposal
- **refine [dimension]** — Further refine specific dimension(s)
- **modify** — Manually edit the idea, then re-evaluate
- **pivot** — Substantially change direction, restart from Phase 1
- **abort** — Stop here, keep all artifacts

Status Determination

• Ready: ALL dimensions ≥ threshold
• Not Ready: Any dimension < threshold after max rounds

Response Handling

• proceed → Phase 5 (works even if "Not Ready" — user's choice)
• refine X → Additional targeted refinement round on dimension X, then re-present
• modify → User provides edited brief, run 1 round of full re-evaluation
• pivot → Return to Phase 1 with new/revised idea
• abort → Save all artifacts, end skill

---

Phase 5: Proposal Generation

Launch two parallel agents via Agent tool. Both receive:

• Refined idea (structured brief, latest version)
• Literature context (all papers from Phase 2 + any Phase 3 supplements)
• Evaluation history (all rounds, scores, refinements, structural limitations)

Agent C: Method Design Writer

Agent C generates the following sections for proposal.md:

1. Title & Abstract

• Concise, venue-appropriate title
• 150-250 word abstract covering problem, approach, key innovation, expected results

2. Motivation & Problem Statement

• Gap identification with real-world relevance
• Formal problem definition (mathematical where appropriate)
• Why existing approaches are insufficient (cite Phase 2 literature)

3. Related Work

• Organized by themes from Phase 2 literature search
• 3-6 thematic subsections
• Positioning statements: "Unlike [X] which..., our approach..."
• Citations as [Author et al., Year]

4. Method Design

• Architecture overview: Mermaid diagram showing system components and data flow

  graph TD
      A[Input] --> B[Component 1]
      B --> C[Component 2]
      C --> D[Output]
  

• Component descriptions: Each component with:
  • Purpose and intuition
  • Mathematical formulation (LaTeX equations: $\mathcal{L} = ...$)
  • Input/output specification
• Algorithm pseudocode:

  Algorithm 1: [Name]
  Input: ...
  Output: ...
  1: for each ... do
  2:   compute ...
  3: end for
  4: return ...
  

• Complexity analysis: Time and space complexity
• Token cost estimation (for agent-based proposals):

  Component	Model	Calls/Run	Input Tokens	Output Tokens	Cost/Run
  Mark as [ESTIMATED]. Include sensitivity note: "Cost is 2-3x if errors require re-runs."

5. Limitations & Risks

• Reference structural limitations from evaluation
• Acknowledge known weaknesses transparently
• Discuss failure modes and mitigation strategies

Agent D: Experiment Design Writer

Agent D generates the experiment sections for proposal.md and a separate experiment-benchmarks.md:

6. Task Definition

• Formal problem setup
• Input/output specification
• Evaluation criteria

7. Benchmark Survey

Search literature for benchmarks on this task:

• Temporal scope: last 5 years for recent, 10-year window for foundational
• Venues: top-tier (NeurIPS, ICML, CVPR, ACL) + domain-specific
• If niche: identify related task benchmarks; note "proposed new benchmark" as contribution

python "BASE_DIR/scripts/paper_search.py" \
  --query "<task> benchmark dataset" \
  --max-results 15 \
  --sort citations

Output table in experiment-benchmarks.md:

Benchmark	Size	Task	Metrics	Used By (Papers)

8. Baseline Survey

Identify SOTA methods:

Method	Year	Venue	Key Result	Code Available?

9. Main Experiment Settings

• Datasets chosen with justification (reference benchmark survey)
• Evaluation metrics (primary + secondary)
• Hyperparameter ranges
• Hardware requirements (GPU type, training time estimate)
• Number of runs / random seeds for statistical validity

10. Ablation Studies

One per key design choice:

Ablation	Component Removed/Replaced	Hypothesis	Metric
w/o Component A	Remove A, keep rest	Performance drops by ~X%	Main metric
Replace B with baseline	Swap B for standard approach	Our B adds ~Y% improvement	Main metric

11. Supplementary Experiments

• Efficiency analysis: Latency, memory, token cost vs. baselines
• Scaling behavior: Vary data size / model size / agent count
• Robustness tests: Noisy input, distribution shift, adversarial examples

12. Qualitative Analysis Plan

• Case studies: 2-3 success cases + 1-2 failure cases with analysis
• Visualization plan: attention maps, agent traces, t-SNE embeddings, decision trees (as appropriate)

13. Quantitative Bonus Analysis

• Statistical significance: paired t-test or bootstrap confidence intervals
• Cross-domain transfer evaluation (if applicable)
• Error analysis: categorize failure types, quantify each category

Assembly

After both agents complete:

1. Merge Agent C and Agent D outputs into single proposal.md
2. Ensure consistent notation (variable names, acronyms) across all sections
3. Add cross-references between method and experiment sections
4. Compile all citations into references.bib:

   python "BASE_DIR/scripts/bibtex_utils.py" \
     merge \
     --output "<output-dir>/references.bib"
   

5. Verify every [Author et al., Year] in proposal has a matching BibTeX entry
6. For non-academic sources (blog posts, preprints): auto-generate minimal BibTeX
7. Save all files to output directory

---

Error Handling

Agent Failures

• Agent A (Literature) fails: Retry with 3 queries. If still fails, proceed with empty context; flag evaluation as "Preliminary (Literature Search Failed)"
• Agent B (Evaluation) fails: Halt and inform user; suggest re-invocation
• Agent C (Method Writer) fails: Fill available sections, mark missing as [TODO]
• Agent D (Experiment Writer) fails: Same — partial output with [TODO] markers
• On partial failure: flag proposal as "Incomplete — manual review required"

Checkpointing

• Each phase saves outputs to disk immediately upon completion
• If session is interrupted, user can re-invoke and reference existing output directory
• idea-evolution.md serves as the checkpoint log

Multi-Idea Input

If Phase 1 detects multiple distinct ideas:

• Present options: evaluate all (parallel), select one, or merge
• Default: prompt user for clarification

---

Usage Examples

Free text idea:

/idea-to-proposal Use graph attention networks to model inter-agent communication in cooperative MARL, where attention weights represent communication bandwidth allocation --target-venue NeurIPS

Paper extension:

/idea-to-proposal https://arxiv.org/abs/2305.12345 --target-venue ICML

(User describes extension in follow-up message after Phase 1 structuring)

From Obsidian note:

/idea-to-proposal ~/Library/Mobile Documents/iCloud~md~obsidian/Documents/My_note/00_Inbox/my-research-idea.md

With custom threshold:

/idea-to-proposal "Diffusion-based program synthesis with formal verification feedback" --threshold 8 --max-rounds 5

---

Dependencies

• paper_search.py — Unified academic paper search (arXiv + Semantic Scholar + DBLP)
• bibtex_utils.py — BibTeX reference management (fetch, merge, validate)
• Claude Agent tool — For launching parallel sub-agents in Phases 2 and 5
• Claude Read tool — PDF/file extraction
• Claude WebFetch tool — URL content extraction

---

Version: 1.0 Last Updated: 2026-03-23