Prose Writer (Evidence-first)

Evidence Steward (skeptic)

Mission: prevent hollow writing by refusing to pad when evidence is thin or underspecified.

Do:

• Preflight each H3 with 4 lines (tension/contrast/eval/limitation) before drafting.
• Keep quantitative claims scoped (task + metric + constraint in the same sentence).
• Stop and route upstream when you would have to guess.

Avoid:

• Strong, unqualified claims when evidence is abstract-only.
• Citation dumps that act like tags rather than evidence.

Coherence Editor (linker)

Mission: make the paper read as a single argument across sections.

Do:

• Weave 1-2 content-bearing transition sentences between adjacent units.
• Ensure Intro/Related Work carries the single evidence-policy paragraph so H3s stay content-focused.

Avoid:

• Planner talk in transitions (semicolons, "setting up a cleaner comparison", "remaining uncertainty is ...").
• Repeating the same discourse stem across many sections (Taken together, In summary, etc.).

Role prompt: Draft Author (evidence-first; paper voice)

Use this as your internal framing while drafting output/DRAFT.md. It is guidance, not a sentence template.

You are writing a technical survey draft from evidence packs.

Your job is to execute argument moves under evidence and citation constraints:
- tension -> contrast -> evaluation anchor -> limitation
- keep every claim attached to citations inside the sentence that needs them
- synthesize across papers (>=2 citations in at least one paragraph per H3)

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

Constraints:
- do not invent facts or citations
- only use citation keys present in citations/ref.bib
- if you cannot write a contrast or evaluation anchor without guessing, stop and route upstream

Non-negotiables

• No prose without approval: for surveys, require Approve C2 in DECISIONS.md.
• No invented citations: only use keys present in citations/ref.bib.
• No placeholder leakage: if any upstream artifact still contains scaffold markers/ellipsis/TODO, do not write; block and request evidence fixes.
• No pipeline voice: do not leak internal scaffolding terms like “working claim”, “enumerate 2-4”, “scope/design space/evaluation practice”.

Writing requirements (explicit contract)

This skill is successful only if the draft reads like an evidence-backed survey, not an outline expansion.

Per-H3 argument requirements (structure)

For each H3 subsection, ensure the prose contains all of the following moves (not necessarily as headings):

• Thesis early: paragraph 1 ends with a clear, conclusion-first thesis sentence (no narration openers).
• Contrast: at least two explicit A-vs-B contrasts (use contrast words; do not write one paragraph per paper).
• Evaluation anchoring: at least one paragraph that states a benchmark/dataset/metric/protocol (and constraints like budget/tool access when relevant).
• Cross-paper synthesis: at least one paragraph with >=2 citations in the same paragraph that explains a pattern/trade-off.
• Limitation: at least one explicit caveat tied to evidence granularity (protocol mismatch, missing ablations, unclear threat model).

Citation requirements (verifiability)

• Use only citation keys in citations/ref.bib.
• Keep citations inside the sentence that carries the claim.
• Avoid citation dumps that act like tags.

Bad:

• Many systems adopt tool schemas. [@a; @b; @c]

Better:

• Systems such as X [@a] and Y [@b] formalize tool schemas to reduce action ambiguity, whereas Z [@c] keeps the interface looser and shifts the burden to validation.

Style requirements (paper voice)

• Do not narrate the outline (avoid: This subsection surveys ..., In this subsection ...).
• Do not use slide navigation (avoid: Next, we move ..., We now turn to ...).
• Put evidence-policy limitations once in front matter; do not repeat "abstract-only" boilerplate across H3s.
• Avoid repeated synthesis stems (e.g., starting many paragraphs with Taken together, ...).

Prevention (before you write)

For each H3, do a short preflight (kept out of the final prose):

• 1 tension sentence
• 1 A-vs-B contrast sentence with >=2 citations
• 1 evaluation-anchor sentence (task/metric/constraint)
• 1 limitation sentence

If you cannot do this without guessing, stop and fix upstream evidence instead of writing filler.

Inputs

• outline/outline.yml
• outline/subsection_briefs.jsonl
• outline/transitions.md
• outline/evidence_drafts.jsonl
• Optional: outline/tables_index.md, outline/tables_appendix.md, outline/timeline.md, outline/figures.md
• Optional: outline/claim_evidence_matrix.md
• citations/ref.bib
• DECISIONS.md

Outputs

• output/DRAFT.md and/or output/SNAPSHOT.md

Decision: snapshot vs draft

• Snapshot: bullets-first, ~1 page; summarize what evidence exists + what is missing.
• Draft: section-by-section prose that follows each subsection’s paragraph_plan and uses paragraph-level citations.

Workflow (v3: planner↔writer, section-by-section)

Before writing, load the structural and coherence inputs: outline/outline.yml (section order) and outline/transitions.md (transition map). Optionally consult outline/claim_evidence_matrix.md as an evidence index.

1. Gate check (HITL)

   • Read DECISIONS.md.
   • If Approve C2 is not ticked, write a short request block (what you plan to write + which evidence packs you will rely on), then stop.

2. Input integrity check (fail fast)

   • Read outline/subsection_briefs.jsonl and confirm every H3 has a brief and the following fields are filled and non-placeholder: scope_rule, rq, axes, clusters, paragraph_plan.
   • Read outline/evidence_drafts.jsonl and confirm every H3 has an evidence pack with:
     • blocking_missing empty,
     • evidence_snippets non-empty,
     • concrete_comparisons >= 3.

3. Planner pass (NO PROSE YET)

   • For each H3 subsection, read its brief + evidence pack and decide:
     • Thesis: 1 sentence that is true for this subsection and would be false in other subsections.
     • Two contrasts: 2 sentences of the form “A vs B” where each side is grounded in specific cited works (not “they differ”).
     • One limitation/failure mode: 1 sentence grounded in the evidence pack’s failures_limitations or snippet provenance.
     • Cite placement: which citations will appear in which paragraph (so citations are evidence, not decoration).
   • If you cannot do this without guessing, stop and push the gap upstream (strengthen briefs/notes/evidence packs) rather than writing template prose.

4. Writer pass (write per subsection; avoid global dump)

   • Write 6–10 paragraphs per H3 following paragraph_plan (survey-quality default).
   • Aim for ~800–1400 words per H3 (shorter only if the evidence pack is explicitly thin and you mark it as provisional).
   • Keep prose natural, but make every paragraph an argument: claim → cited evidence → synthesis.
   • Evidence policy placement: if the survey is primarily abstract-based, put a single short evidence-policy paragraph once (prefer Introduction or Related Work). Avoid execution-log phrasing like this run .... Do not create a dedicated “Evidence note” heading by default, and do not repeat the same evidence-mode disclaimer sentence in every H3; only mention verification needs when they are subsection-specific.
   • Enforce scope_rule strictly to prevent silent drift; if you include an out-of-scope paper as a bridge, justify it once and keep it secondary.

5. Weave transitions (coherence)

   • Between adjacent subsections/sections, add 1–2 transition sentences that reflect the taxonomy logic (not generic “Moreover/However”).

6. Integrate cross-cutting artifacts (paper-like)

   • Tables are part of the default survey deliverable. If outline/tables_appendix.md exists, place its contents into the draft as an Appendix block (recommended: after Conclusion). Do not paste outline/tables_index.md into the paper; it is an internal index.
   • outline/timeline.md and outline/figures.md remain optional/intermediate by default: weave them into relevant prose (or a short appendix) only if they add real reader value.
   • Prefer referencing tables in prose over restating an identical “axes list” sentence in every subsection.

7. Self-check + revise (hard fail signals)

   • If the draft contains ..., unicode ellipsis …, scaffold phrases (e.g., “enumerate 2-4 …”), or repeated boilerplate sentences, treat it as a pipeline failure signal and rewrite.
   • If tables contain truncation or instruction-like text, regenerate them upstream (C4) rather than patching them into the prose.

Anti-template smells (rewrite if repeated)

These phrase families are a strong “generator voice” signal. If they appear, rewrite them into content claims (or delete) without adding new facts/citations:

• “Scope and definitions … / Design space … / Evaluation practice …”
• “enumerate 2-4 …”
• “We use the following working claim …”
• “Across representative works, the dominant trade-offs …”
• “A useful way to compare approaches is …”
• “abstracts are treated as verification targets …”
• “The main axes we track are …”
• “This subsection surveys/argues …” / “In this subsection …”
• Slide navigation: “Next, we move from …” / “We now turn to …”
• Injection-like enumerators: “A few representative references include …” / “Notable lines of work include …” / “Concrete examples ... include ...”
• Meta process advice: survey synthesis/comparisons should ...
• Repeated synthesis openers (e.g., Taken together, ... at the start of many paragraphs)
• Repeated opener labels across many subsections (e.g., literal Key takeaway:)

Quality checklist

• No …, TODO, (placeholder), or <!-- SCAFFOLD --> remains in output/DRAFT.md.
• Every subsection has citations and at least one paragraph with >=2 citations (cross-paper synthesis).
• No undefined citation keys (all keys exist in citations/ref.bib).
• Scope is consistent with GOAL.md and scope_rule.
• Subsections are not thin (avoid 2-paragraph ~150-word stubs; expand using evidence packs).

Helper script (bootstrap)

The helper script is a gate wrapper: it blocks until approvals + prerequisites are satisfied and a real output/DRAFT.md exists (no scaffold markers). Writing itself is LLM-driven.

Quick Start

• python .codex/skills/prose-writer/scripts/run.py --help
• python .codex/skills/prose-writer/scripts/run.py --workspace <workspace_dir>

All Options

• See --help.

Examples

• Run the gate wrapper after approval (it will block until output/DRAFT.md is written):
  • Tick Approve C2 in DECISIONS.md then run:
  • python .codex/skills/prose-writer/scripts/run.py --workspace workspaces/<ws>

Troubleshooting

Issue: writer outputs ellipsis / scaffold text

Symptom: output/DRAFT.md contains …, enumerate 2-4 ..., or repeats the same paragraph template.

Causes:

• outline/subsection_briefs.jsonl is missing or generic.
• outline/evidence_drafts.jsonl has blocking_missing or scaffold markers.

Solutions:

• Fix upstream: regenerate briefs/evidence packs, enrich abstracts/fulltext, and block writing until evidence is concrete.

Issue: scope drift (e.g., T2I vs T2V)

Symptom: subsections cite many out-of-scope papers without justification.

Solutions:

• Tighten scope_rule in subsection briefs and rerun evidence packs.
• Tighten queries.md excludes and rerun retrieval/dedupe/mapping.