Stage 6: Writing

The writing stage produces the paper draft. The Writer agent works in clean, isolated sessions, receiving curated context from the Orchestrator for each section.

Entering This Stage

What you have:

• Complete analysis with paper framing (experiments/exp-001/analysis.md)
• All figures generated (papers/figures/)
• Related work summaries from Scout (papers/related_work/)
• Paper outline from Planner (papers/outline.md)

What you don't have yet:

• Paper text
• LaTeX document
• Final bibliography

Steps

graph TD
    A[1. Outline Refinement] --> B[2. Related Work Prep]
    B --> C[3. Section Writing]
    C --> D[4. Integration]
    D --> E[5. Internal Review]
    E --> F{Quality OK?}
    F -->|Yes| G[Gate: Advance to Review]
    F -->|No| C

    style A fill:#f9f0ff,stroke:#7c3aed
    style B fill:#ecfdf5,stroke:#059669
    style C fill:#ede9fe,stroke:#7c3aed
    style E fill:#fef3c7,stroke:#d97706

1. Outline Refinement

Agent: Orchestrator (Claude Opus) with Planner

The Orchestrator refines the paper outline based on actual results (which may differ from the original plan):

# papers/outline.md (refined)
## 1. Introduction (1 page)
- Problem: Attention is O(n^2), limiting long-sequence modeling
- Key insight: Flash attention tiling + recurrent formulation
- Contributions (3 bullets)

## 2. Related Work (1 page)
- Efficient attention methods
- Recurrent approaches to attention
- Hardware-aware optimization

## 3. Method (2 pages)
- 3.1 Preliminaries: Retention mechanism
- 3.2 Flash-Recurrent Attention
- 3.3 Analysis: Complexity and memory

## 4. Experiments (2 pages)
- 4.1 Setup
- 4.2 Main Results (Table 1)
- 4.3 Ablation Study (Table 2)
- 4.4 Scaling Analysis (Figure 2)

## 5. Discussion (0.5 pages)
- Limitations
- Future work

## 6. Conclusion (0.5 pages)

2. Related Work Preparation

Agent: Scout (Gemini)

The Scout ensures all cited papers have proper summaries and BibTeX entries:

• Verify all baselines have entries
• Add any papers discovered during analysis
• Cross-check with analysis.md for completeness

Output: Updated papers/related_work/summaries.yaml and papers/related_work/papers.bib

3. Section Writing

Agent: Writer (Claude Opus, clean session per section)

The core writing phase. Each major section gets a fresh Writer session:

Section	Context Provided	Key Challenge
Introduction	Outline, contributions, related work overview	Motivating the problem clearly
Related Work	Pre-digested summaries, positioning notes	Connecting prior work to ours
Method	Algorithm description, pseudocode, complexity analysis	Technical precision + clarity
Experiments	Results tables, figures, statistical tests	Honest presentation of results
Discussion	Analysis synthesis, known limitations	Balanced assessment
Conclusion	Full paper context (if space allows)	Concise, forward-looking
Abstract	Full paper (written last)	Compress everything into 200 words

Sections are written in dependency order

Introduction and Method can be written in parallel. Experiments depends on having the Method section (for consistent notation). Discussion depends on Experiments. Abstract is written last.

graph LR
    I[Introduction] --> E[Experiments]
    M[Method] --> E
    RW[Related Work] --> E
    E --> D[Discussion]
    D --> C[Conclusion]
    C --> A[Abstract]

    style A fill:#fef3c7,stroke:#d97706

4. Integration

Agent: Orchestrator (Claude Opus)

The Orchestrator integrates all sections into a single LaTeX document:

• Verify notation consistency across sections
• Check cross-references (figures, tables, equations)
• Ensure page limit is respected
• Compile LaTeX to verify it builds

Output: papers/drafts/v1.tex (complete first draft)

Integration is not just concatenation

The Orchestrator reads through the integrated draft looking for inconsistencies: a variable named differently in Method vs. Experiments, a claim in the Introduction not supported in Experiments, a related work paper cited but not discussed.

5. Internal Review

Agent: Orchestrator (Claude Opus)

Before the formal review stage, the Orchestrator does a self-check:

• Does the paper tell a coherent story?
• Are all claims supported by data?
• Is the contribution clearly stated?
• Are limitations acknowledged?

If issues are found, specific sections are re-dispatched to the Writer with revision instructions.

Gate

Gate Type	Recommended	Behavior
human	Yes	User reads the draft before formal review
auto-judge	Possible	Judge does a quick quality check
auto	Not recommended	Paper quality deserves review

Read your paper before review

The writing stage gate is a critical human checkpoint. Reading your own paper draft takes 30-60 minutes and often catches issues that no automated review can find — you know your story better than any LLM.

Error Handling

Error	Recovery
Writer context fills mid-section	Start new session, provide partial draft
Section too long for page limit	Orchestrator gives Writer a word budget
LaTeX compilation error	Coder fixes LaTeX syntax
Notation inconsistency	Re-dispatch affected section to Writer
Missing citation	Scout fetches paper, Writer integrates

Writer Context Management

The Writer's clean context is critical for quality. Here's how it works per section:

Session starts (100% context available)
├── Outline loaded (2%)
├── Section spec loaded (3%)
├── Relevant results loaded (10%)
├── Related work summaries loaded (5%)
├── Style guidelines loaded (2%)
├── [78% available for writing]
└── Section written → saved to disk
Session ends

What if 78% isn't enough?

For very long sections (e.g., a 4-page experiments section), the Orchestrator splits it into subsections. Each subsection gets a fresh Writer session. The integration step then ensures coherence across subsections.

Outputs Summary

File	Contents
papers/drafts/v1.tex	Complete first draft
papers/drafts/current.tex	Symlink to latest version
papers/related_work/papers.bib	Complete bibliography
papers/figures/	All figures referenced in paper

Next Stage

When the gate passes, the pipeline advances to Review with a complete paper draft.