subsection-writer_skill

Overview

This skill drafts paper-style subsections as small, inspectable files under sections/ so each unit can be QA'd independently before merge. It enforces evidence-bounded argument moves (tension → contrast → evaluation → limitation) and produces a sections manifest that downstream gates can validate.

How this skill works

The skill reads writer and evidence packs to determine citation scope and required files, then writes H3 bodies and chapter leads as body-only markdown files under sections/. It relies on a manifest JSONL that records required files, existence, and optional citation summaries so writer-selfloop can pass/fail with actionable TODOs. If evidence is insufficient, the flow routes upstream instead of producing filler text.

When to use it

• After DECISIONS.md includes the Approve C2 token and writer packs exist
• When you want evidence-bounded drafting instead of one-shot monolithic drafts
• To break a chapter into reviewable H3 files so reviewers can QA each claim independently
• When citation scope must be strictly enforced per subsection
• When you need a machine-readable sections_manifest.jsonl for gating

Best practices

• Draft each H3 as a content-first argument: open with a tension and end paragraph 1 with a thesis
• Preflight every H3 with a 4-line plan (tension, A-vs-B with citations, evaluation anchor, limitation) and route upstream if you cannot fill it
• Keep H3 files body-only (no headings) and keep chapter leads short and connective
• Validate citation keys against citations/ref.bib and keep citations subsection- or chapter-scoped
• Use the manifest script to generate sections/sections_manifest.jsonl so writer-selfloop can give precise failures

Example use cases

• Produce H3 drafts for a literature survey where each subsection will be reviewed independently
• Create chapter leads that set the lens and connect H3s without repeating pack wording
• Gate a writing sprint: generate small files and iterate with writer-selfloop until all files pass
• Repair weak sections by routing thin packs to evidence-selfloop instead of writing unsupported claims
• Generate a manifest after manual edits so downstream validators see up-to-date existence and citation summaries