Drafting

Drive prose production for the LaTeX manuscript. For every section in .writing/plan.md, resolve evidence for every claim first, then write % claim: id-tagged paragraphs (LaTeX line comments) in manuscript/*.tex. This skill owns the writing-specific prompt template, the Zotero-aware evidence loop, and the graphical abstract dispatch; orchestration is handled by a Claude Code dynamic workflow (parallel section drafting + a two-stage review pipeline) or, as a manual fallback, by superpower-writing:executing-plans.

Announce at start: "I'm using the drafting skill to produce manuscript prose."

Overview

Drafting is the stage where claims become sentences. The hard rule is claim-first: no prose may cite a claim whose STATUS is still stub. Resolve evidence, then write.

Claim-first protocol: see superpower-writing:main §Claim-First Protocol.

Build the terminology ledger before any section drafts. A paper must use one name for one thing; a term that drifts across sections reads as careless work and erodes reviewer trust. Before dispatching the first section, read skills/_shared/core/terminology-ledger.md and extract every recurring load-bearing term (systems, models, datasets, metrics, notation, acronyms, named contributions) into the canonical-term table it specifies. Present the table to the user, flag every collision, and lock the canonical forms before prose starts. The ledger is the single source of truth every section drafter consults instead of re-deciding a name per section. It is advisory: surface collisions and let the user pick the canonical form, never coin a name to fill a gap. When the project opts into .writing/glossary.md, materialize each load-bearing ledger row into one glossary entry so the enforce-terms.py gate can check define-before-use ordering.

This skill only shapes the per-section prompt and the bookkeeping. Orchestration is picked by mode:

• workflow → a Claude Code dynamic workflow drafts independent sections in parallel with the section-drafter agent, then runs the two-stage review as a pipeline (spec-reviewer for outline/claim alignment, then manuscript-reviewer for writing quality). Default for multi-section papers.
• manual-batch → executing-plans (batch execution across a separate or manual session with checkpoints). Fallback when workflows are unavailable or the user wants explicit per-batch review.