Paper Outline

Workflow

Step 0: Load Workspace Context

Before generating the outline:

1. Call workspace get_state to confirm the workspace is initialised.
2. Call workspace read_knowledge for all knowledge files (landscape.md, gap-analysis.md, related-work.md).
3. Retrieve the validated idea from idea_store (action: get, including the methodology field).

This grounds the outline in the accumulated understanding and avoids inventing related work.

Step 1: Context Ingestion

Collect and structure:

• Idea: problem statement + core approach
• Experimental plan: method, baselines, datasets, metrics
• Related papers: grouped by sub-topic
• Key claims: what the paper proves (1 main claim + 2–3 supporting claims)

Step 2: Title and Framing (Self — Orchestrator)

Generate 3 candidate titles following these patterns:

• Descriptive: "Method Name: Improving X by Y via Z"
• Problem-framing: "Why Does X Fail? Insights and a Solution via Y"
• Results-driven: "X% Better Y with Simple Z"

Select the best one with brief justification.

Step 3: Outline Generation (Writer)

Launch @writer with the full context. The writer produces:

Paper: [Title]

Abstract (draft):
[150–250 words covering: motivation → problem → approach → key result → significance]

Section 1: Introduction
  - Hook: [what readers care about]
  - Problem statement: [precise formulation with {formal notation placeholder}]
  - Our approach: [brief preview]
  - Contributions: [3–5 bullet points — be specific: "We prove X", "We show Y improves Z by W%"]
  - Organization: [one sentence]
  Estimated length: 1.5–2 pages

Section 2: Related Work  
  - [Subsection for each related area, 2–4 papers each]
  - Position statement: [explicit contrast with our approach for each subsection]
  Estimated length: 1–1.5 pages

Section 3: Method / Approach / Model
  - 3.1 Problem Formulation: [{formal notation} — define all variables]
  - 3.2 [Core component 1]: [description + figure placeholder]
  - 3.3 [Core component 2]: [description + equation placeholders]
  - 3.4 Theoretical Analysis: [if applicable — key theorems/proofs to include]
  Estimated length: 2–3 pages

Section 4: Experiments
  - 4.1 Experimental Setup: [datasets, baselines, metrics, hardware, hyperparams]
  - 4.2 Main Results: [table placeholder — rows × columns × expected numbers]
  - 4.3 Ablation Studies: [ablation table placeholder]
  - 4.4 Analysis and Discussion: [key findings to highlight]
  Estimated length: 2.5–3 pages

Section 5: Conclusion
  - Summary: [2 sentences on what we showed]
  - Limitations: [3 honest limitations — required by NeurIPS checklist]
  - Future Work: [3 natural extensions]
  Estimated length: 0.5 pages

References: [estimated 25–40 citations]

Total estimated page count: 8–10 pages (NeurIPS/ICML main track)

Step 4: Section Drafts (Writer)

Optionally draft prose for:

• Introduction (usually best to draft early)
• Abstract (written last but planned first)
• Related Work (can be templated from literature survey)

Launch @writer for each section separately for higher quality.

Step 5: Checklist

Run the NeurIPS reproducibility checklist:

• All claims in the paper are supported by an experiment or proof
• Datasets are publicly available or release plan is described
• Compute requirements are stated
• All hyperparameters are reported
• Code will be released (checkbox required)
• Limitations section exists

Step 6: Save and Output

Update idea in idea_store (action: update):

• Store outline in outline field
• Update status to in_progress (if not already)

Produce a final markdown document:

# Paper Outline: [Title]
[Full structured outline from Step 3]

## Section Drafts
[If requested]

## NeurIPS Checklist
[From Step 5]

Quality Criteria

• Title must avoid hyperbole ("novel", "revolutionary", "unprecedented")
• Contributions must be specific and verifiable claims (not vague improvements)
• Abstract must mention the key quantitative result (even if estimated)
• Limitations section must be honest (not "limited to supervised learning" as the only limitation)
• Introduction must contrast with the 3 most related papers explicitly