yzxoi/paperly-codex/draft-generation

Draft Generation

Overview

Generates structured manuscript drafts — or specific sections — grounded in the available literature base. Unlike generic AI writing, this skill is intended to stay citation-aware, claim-traced, and calibrated to the evidence grades of the underlying papers.

When to Use

• User asks "write a draft introduction for our paper on [topic]"
• User wants a related work section that synthesizes, not just lists
• User needs a first draft of a methods section based on prior work
• User asks "turn our living review into a manuscript draft"
• Lab wants to generate a first pass at a paper to react to and refine

Key Capabilities

• Generate full paper sections: introduction, related work, methods, discussion
• Ground all claims in the available literature corpus — no hallucinated citations
• Calibrate claim confidence to evidence grades from evidence-grading
• Apply hedging language automatically for preliminary findings
• Match target venue style (NeurIPS, Nature, PNAS, ICML, etc.)
• Insert placeholder tags for figures, tables, and results not yet available
• Generate an outline first and await approval before expanding

Usage Examples

Generate a related work section

draft_generation.related_work(
    topic="contrastive learning for protein representation",
    corpus=review.get_papers(),
    target_venue="NeurIPS",
    max_words=800,
    style="synthesizing"  # synthesizing | chronological | thematic
)

Generate introduction from living review

draft_generation.introduction(
    living_review=review.export(),
    research_question="Can self-supervised protein LMs generalize to low-data fitness prediction?",
    target_venue="Nature Methods",
    max_words=600
)

Generate full paper outline

draft_generation.outline(
    title="Proposed title here",
    abstract="Proposed abstract here",
    target_venue="ICML",
    await_approval=True
)

Output Format

All output is structured Markdown with inline citations in the format [AuthorYear]. A reference list is appended. Placeholder tags use [FIGURE: description] and [RESULT: metric] syntax. Evidence grade annotations included in a review mode.

Best Paired With

• paper_editor after the review agents have already identified what should change
• writing_reflector when a section needs both structural repair and rewriting

Notes

• All citations should be verified against the available corpus — no fabricated references
• Always review AI-generated drafts carefully: structure and flow are strong, factual nuance requires human review
• Combine with cross-paper-synthesis for the related work section and gap-detection for the motivation or novelty framing