Subsection Writer (compatibility router)

Triggers & routing

• Trigger: subsection writer, per-section writing, split sections, sections/, 分小节写, 按章节拆分写作.
• Use when: Approve C2 is recorded and writer packs exist (outline/writercontextpacks.jsonl); you want evidence-bounded drafting without a monolithic one-shot draft.

Purpose: write or refine per-section survey prose under sections/ while keeping the current pipeline contract unchanged.

Compatibility mode:

• output paths stay the same
• scripts/run.py still handles approval checks, missing-file bootstrap, and sections/sections_manifest.jsonl
• writing guidance now lives in references/ instead of being encoded primarily in the script

Load Order

Always read:

• references/overview.md
• references/paragraph_jobs.md
• references/paragraphjobarchetypes.md when adjusting paragraph-role behavior or refactoring writer policy out of Python
• references/bootstrap_assembly.md when reasoning about compatibility-mode bootstrap behavior

Read as needed:

• references/opener_catalog.md when paragraph 1 sounds generic or narrated
• references/contrast_moves.md when building A-vs-B comparison paragraphs
• references/evalanchorpatterns.md when making performance / robustness / benchmark claims
• references/limitation_moves.md when adding caveats or local conclusions
• references/examplesgood.md and references/examplesbad.md for calibration only

Machine-readable contract:

• assets/subsectionwritercontext.schema.json
• assets/bootstrapparagraphtemplates.json
• assets/paragraphjobtemplates.json

Inputs

Required:

• DECISIONS.md (must include Approve C2)
• outline/outline.yml
• outline/writercontextpacks.jsonl (preferred)
• citations/ref.bib

Optional but useful:

• outline/subsection_briefs.jsonl
• outline/evidence_drafts.jsonl
• outline/evidence_bindings.jsonl
• outline/anchor_sheet.jsonl
• outline/chapter_briefs.jsonl

Outputs

Keep the current contract:

• sections/S<sub_id>.md for H3 bodies
• sections/sections_manifest.jsonl, refreshed across all section files already

produced by the three writer Skills

• sections/h3_bodies.refined.ok after a model or human has reviewed the

generated H3 bodies and is submitting U100 for mandatory acceptance

front-matter-writer owns Abstract, Introduction, Related Work, Discussion, and Conclusion. chapter-lead-writer owns H2 lead blocks. This Skill reads those files while refreshing the shared manifest but does not claim their authorship.

Writer policy

The active rule is move coverage, not paragraph quota. A subsection should cover the necessary argument moves the pack supports; do not pad to a fixed count when the evidence does not justify it.

Opener / ending policy:

• generate 2-4 opener candidates from the pack's actual tension, contrast, protocol, or limitation signals; keep the most content-bearing option instead of reusing one stock stem everywhere
• let subsection endings emerge from evidence-bearing comparison / limitation material; do not append a fixed “safest synthesis / decision rule” closer just to make the paragraph feel finished
• normalize internal axis labels into natural reader prose; slash-style brief handles should remain planning metadata, not leak unchanged into the paper

Script boundary

Use scripts/run.py as a helper only:

• it may bootstrap missing H3 files and refresh the manifest
• bootstrap prose is assembled deterministically from writer packs and the

versioned template assets; it is inspectable fallback material, not a claim of model authorship or final prose quality

• it must not be treated as the canonical source of prose shape or voice policy
• it must not self-certify its bootstrap prose: the script never creates sections/h3_bodies.refined.ok
• the marker is only U100's submission attestation; it is created before the

downstream writer-selfloop, so it cannot mean that the self-loop already passed

• mandatory acceptance independently compares H3 sentences with the Run-selected

writer template assets and rejects literal residue above the Pipeline limit

• writer-selfloop calls the shared strict section checker, which recomputes the

H3 measure before the report can PASS

• the mandatory pipeline-auditor measures the entire merged draft and writes a

scorecard whose verdict and dimensions the Harness projects into the evaluation ledger

• the current 10% limit is an initial policy target; a published current-contract

replay measures 0/226 residue for one retained Artifact set, while clean from-scratch and cross-topic reproduction remain open

• if the marker predates a writer input or the writer script, it is stale and the next run removes it before regenerating bootstrap prose

Quick Start

• uv run python .codex/skills/subsection-writer/scripts/run.py --workspace <workspace>

Troubleshooting

• If the pack is thin, stop and route upstream instead of padding prose.
• If the subsection sounds narrated, reload references/openercatalog.md and references/examplesbad.md.
• If a claim lacks protocol context, reload references/evalanchorpatterns.md before rewriting.

Execution notes

When running in compatibility mode, scripts/run.py currently consumes:

• DECISIONS.md for Approve C2
• outline/outline.yml to enumerate chapter / subsection files
• outline/writercontextpacks.jsonl as the primary drafting input
• citations/ref.bib for in-scope citations
• outline/subsectionbriefs.jsonl, outline/evidencedrafts.jsonl, outline/evidencebindings.jsonl, outline/anchorsheet.jsonl, and outline/chapter_briefs.jsonl as optional enrichment sources

Script

Quick Start

• uv run python .codex/skills/subsection-writer/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir>
• --unit-id <id>
• --inputs <a;b;...>
• --outputs <a;b;...>
• --checkpoint <C*>

Examples

• uv run python .codex/skills/subsection-writer/scripts/run.py --workspace <workspace>

Troubleshooting

• If DECISIONS.md lacks Approve C2, stop and fix approval first.
• If outline/writercontextpacks.jsonl is thin, reroute upstream instead of padding prose.
• If citation scope looks wrong, re-check citations/ref.bib and the writer packs before editing output text.