Subsection Writer

Subsection Writer (section author playbook)

Purpose: produce a paper-like draft by writing small, inspectable files under sections/.

This skill is semantic:

• you are not "generating text"
• you are executing argument coverage (thesis + contrasts + protocol/evaluation anchoring + limitations) under a citation-scope contract
• you avoid a fixed paragraph macro; the required moves must appear, but order and paragraph shapes should vary across H3s

Use writer-selfloop as the strict gate (PASS/FAIL + actionable TODO). Use evidence-selfloop to route upstream when packs are too thin.

Default mode: manual (LLM-first)

Treat this skill as a writing playbook first, and as an automation wrapper second.

• The core work is writing sections/*.md in paper voice under the writer-pack / citation-scope contract.
• Any scripts (if present) must remain non-prose helpers (manifest generation, validation). They are optional.

Role cards (use explicitly)

Section Author (content expert)

Mission: draft H3 bodies that execute argument moves under citation scope.

Do:

• Choose an opener mode (tension/contrast/protocol/failure/decision/lens) and end paragraph 1 with a thesis.
• Make explicit A-vs-B contrasts grounded in in-scope citations.
• Include a protocol-aware evaluation anchor (task/metric/constraint).
• State at least one subsection-specific limitation.

Avoid:

• Narration openers and slide navigation.
• Copying axis labels from briefs/packs into prose.

Evidence Steward (skeptic)

Mission: refuse to pad when the pack cannot support concrete contrasts.

Do:

• Preflight each H3 with 4 lines (tension/contrast/eval/limitation).
• Route thin-evidence sections upstream (evidence-selfloop) instead of writing filler.

Avoid:

• Strong claims without protocol context.

Style Harmonizer (paper voice)

Mission: keep tone calm and paper-like while drafting.

Do:

• Vary openers; keep signposting light and content-bearing.
• Keep evidence-policy disclaimers out of H3 bodies (front matter once).

Avoid:

• Repeating template stems (e.g., Taken together, This subsection...) across many sections.
• Internal shorthand that reads like planning notes once merged (e.g., calling assumptions "tokens"); prefer "protocol details/assumptions/metadata".

Role prompt: Section Author (content expert)

Use this as your internal system-style framing while drafting H3 bodies (sections/S<sub_id>.md). It is guidance, not a sentence template.

You are a section author for a technical survey.

Your job is not to “generate text”, but to execute argument moves under evidence and citation constraints:
- state a concrete tension and end paragraph 1 with a thesis
- make explicit A-vs-B contrasts grounded in in-scope citations
- anchor at least one claim in an evaluation protocol (task/metric/constraint)
- synthesize across papers (>=2 citations in the same paragraph)
- state a limitation that changes interpretation (protocol mismatch, unclear threat model, missing ablations)

Style:
- calm, academic, content-bearing
- no outline narration (“This subsection…”) and no slide navigation (“Next, we…”)
- no pipeline jargon (workspace/unit/evidence pack/quality gate)

Constraints:
- do not invent facts or citations
- only use citation keys present in citations/ref.bib
- keep citations within the allowed scope for this H3 (evidence bindings / writer pack)

Tip: writer packs already carry role cards (section_author / evidence_steward / style_harmonizer) copied from .codex/skills/writer-context-pack/assets/paper_voice_palette.json. Switch roles deliberately: draft as section_author, sanity-check as evidence_steward, then smooth rhythm as style_harmonizer.

Inputs

• DECISIONS.md (must include Approve C2)
• outline/outline.yml (ordering + ids)
• outline/chapter_briefs.jsonl (H2 throughlines)
• outline/writer_context_packs.jsonl (preferred drafting pack per H3)
• outline/evidence_bindings.jsonl (citation scope)
• citations/ref.bib

Optional (improves specificity and prevents hollow prose):

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

Outputs (what must exist before merge)

Global files:

• sections/abstract.md (MUST include ## Abstract or ## 摘要)
• sections/discussion.md (must include ## Discussion)
• sections/conclusion.md (must include ## Conclusion)

Per-outline files:

• For each H2 section without H3 subsections: sections/S<sec_id>.md (body-only; no headings)
• For each H2 section with H3 subsections: sections/S<sec_id>_lead.md (body-only; no headings)
• For each H3 subsection: sections/S<sub_id>.md (body-only; no headings)

Helper artifact (generated by the script; used by writer-selfloop):

• sections/sections_manifest.jsonl

sections/sections_manifest.jsonl (contract, not just a script output)

This manifest is how later gates stay debuggable without "reading your mind". If you do not use the helper script, write it manually as JSONL.

Required record fields (per line):

• kind: global | h2 | h2_lead | h3
• id, title (section/subsection id + title from outline/outline.yml)
• section_id, section_title (for h2_lead/h3)
• path (e.g., sections/S3_2.md, sections/S3_lead.md, sections/abstract.md)
• exists (bool; whether the file exists and is non-empty)

Optional (recommended; improves routing and prevents citation drift):

• citations (extracted keys present in the file; informational)
• allowed_bibkeys_selected|mapped|chapter|global (from outline/evidence_bindings.jsonl)
• anchor_facts (small examples from outline/anchor_sheet.jsonl)
• sha1, bytes (for change tracking)

Use specialized playbooks (recommended)

• Front matter (Abstract / Intro / Related Work / Discussion / Conclusion): use front-matter-writer.
• Chapter leads (sections/S<sec_id>_lead.md): use chapter-lead-writer.
• Local cleanup of a single weak H3: use subsection-polisher (no citation key changes).

This skill still owns the end state: all required sections/*.md files exist and read like paper prose.

H3 writing contract (argument moves, not templates)

A150++ minima (survey deliverable):

• =10 paragraphs (substantive, not captions)

• =12 unique citation keys in the H3

• =3 anchored+cited paragraphs (cited + (digit OR evaluation token OR limitation token))

• =2 explicit A-vs-B contrast paragraphs + >=1 multi-cite synthesis paragraph (>=2 cites in one paragraph)

Each H3 file should do the following work (order can vary, but all must appear):

• Thesis early: end paragraph 1 with a clear, conclusion-first takeaway aligned to the pack\x27s thesis.
• Concrete contrasts: compare at least two approaches along one axis at a time (use explicit contrast words: whereas / in contrast / unlike).
• Protocol-aware evaluation: when making performance/robustness claims, include minimal context (task + metric + constraint/budget/tool access) in the same paragraph.
• Cross-paper synthesis: at least one paragraph that synthesizes >=2 cited works into a pattern (not a list).
• Limitations: name at least one limitation that changes how to interpret results (protocol mismatch, unclear threat model, missing ablations).

Paper voice (keep it natural):

• Avoid narration openers ("This subsection surveys...") and slide navigation ("Next, we move...").
• Avoid repeating the same discourse stem across many sections ("Additionally", "This suggests", "Taken together").
• Avoid count-based paragraph openers as a reusable slot ("Two limitations...", "Three takeaways..."); vary phrasing and embed caveats inside contrast paragraphs instead.
• Never leak pipeline words (workspace/unit/evidence pack/quality gate) into the draft.

Write For Curation (avoid 'only gets longer')

This pipeline includes a C5 curation pass (paragraph-curator). Draft H3 paragraphs so they are easy to keep/merge:

• One primary move per paragraph (claim/setup/contrast/eval/boundary/conclusion). End with a reusable output sentence.
• Avoid kitchen-sink paragraphs (many axes + many cites). Curation works by fusing 2-3 clean paragraphs.
• If a paragraph is high-leverage (paragraph-1 thesis sentence, synthesis paragraph, evaluation anchor), draft 2 candidate versions in parallel (different angle), then keep the better one (or fuse two winners). Do not leave both candidates in the final text.
• Stay near the paragraph budget (survey: 10-12, deep: 11-13). If you need more coverage, replace weak paragraphs rather than appending.
• Keep citation keys stable within the subsection (no cross-subsection moves). If you need more citations later, rely on citation-diversifier/citation-injector (in-scope, NO NEW FACTS).

Opener palette (avoid "same first sentence everywhere")

Every H3 needs an opener, but openers must not look like a generated table-of-contents. Pick an opener mode per H3 (the writer pack may already suggest one via opener_mode), and paraphrase.

Allowed opener moves (choose 1; keep it content-bearing):

• Tension-first: start with the subsection trade-off, end paragraph 1 with the thesis.
• Decision-first: frame the choice system builders face (what to optimize, what breaks).
• Failure-first: start from a concrete failure mode that motivates the lens.
• Protocol-first: start from comparability constraints (what makes results interpretable).
• Contrast-first: open with an A-vs-B sentence (then explain why it matters).
• Lens-first: open with the chapter lens, then narrow to this subsection question.

High-signal anti-patterns (rewrite immediately):

• This subsection surveys/argues ...
• In this subsection, we ...
• This section provides an overview ... (keep "overview" narration rare; prefer tension/lens openers)
• Next, we move ... / We now turn ...
• Count-based openers used as a slot: Two limitations... / Three takeaways...

Preflight (before you draft an H3)

Write (privately) a 4-line plan for the subsection:

• Tension sentence (what trade-off is real here)
• A-vs-B contrast sentence (>=2 citations embedded)
• Evaluation anchor sentence (task/metric/constraint)
• Limitation sentence (what would falsify / when it doesn\x27t transfer)

If you cannot write this plan without guessing, stop and route upstream (run evidence-selfloop and fix packs) instead of writing filler. If the opener starts sounding slot-like, draft the middle paragraphs first (contrasts + protocol anchors + limitations), then rewrite paragraph 1 last so the opener reflects the section content.

Workflow (minimal, high leverage)

1. Gate on approval

• Confirm DECISIONS.md includes Approve C2.

2. Enumerate required files

• Use outline/outline.yml to list every H2/H3 unit and its required sections/*.md file.

3. Load chapter throughlines

• Use outline/chapter_briefs.jsonl to write each sections/S<sec_id>_lead.md in a way that previews the lens and connects the H3s.

4. Draft H3 bodies from packs

• Use outline/writer_context_packs.jsonl as the source of truth for each H3.
• Use outline/evidence_bindings.jsonl to stay in citation scope.
• Use outline/evidence_drafts.jsonl and outline/anchor_sheet.jsonl (if present) to force concrete contrasts, protocol anchors, and limitation hooks.
• If a pack is missing/thin, fall back to outline/subsection_briefs.jsonl for scope/thesis/paragraph_plan (do not copy its phrasing verbatim).
• Validate citation keys against citations/ref.bib as you write.

5. Iterate with the gate

• Run writer-selfloop and fix only the failing files it lists.

Mini rewrite recipes (paraphrase; don\x27t copy)

Narration opener -> content claim:

• Bad: This subsection surveys tool interfaces for agents.
• Better: Tool interfaces define what actions are executable; interface contracts therefore determine which evaluation claims are interpretable across environments.

Meta "survey should" -> literature-facing observation:

• Bad: Therefore, survey comparisons should control for tool access.
• Better: Across reported protocols, tool access and budget assumptions vary widely, which makes head-to-head comparison fragile unless those constraints are normalized.

Count-based limitation slot -> content-bearing caveat (vary across H3s; don't reuse the same opener):

• Bad: Two limitations stand out. First, ...
• Better options (pick one; rotate across H3s):
  • Interpretation depends on ...; this matters because it changes how results transfer across protocols.
  • These results hinge on ...; under different budgets/protocols, the reported gap may not transfer.
  • Evidence is thin when ...; claims should be scoped to the reported protocol.

Script

Optional helper only. It does not write prose. It generates/refreshes sections/sections_manifest.jsonl (expected files + allowed citations + anchor facts), which is used by writer-selfloop.

Quick Start

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

All Options

• --workspace <dir>
• --unit-id <U###>
• --inputs <semicolon-separated>
• --outputs <semicolon-separated>
• --checkpoint <C#>

Examples

• Regenerate the manifest after you add/update sections/*.md:
  • python .codex/skills/subsection-writer/scripts/run.py --workspace workspaces/<ws>