Writer Selfloop Skill for Codex

Writer Self-loop (fix only what fails)

Purpose: make writing converge without rewriting the whole paper.

This is the writing self-healing loop:

Default mode: semantic triage (LLM-first)

Treat the gate as a router, not as a reason to rewrite everything.

• First, classify each failure as: evidence substrate vs writing execution vs voice contamination vs citation hygiene.
• Then, fix the earliest responsible artifact (often not the merged draft).
• Rewrite only the failing files; do not churn working sections.

If you do not run the helper script, you can still execute this skill by reading sections/*.md + packs and writing output/WRITER_SELFLOOP_TODO.md manually in the same PASS/FAIL style.

Role cards (use explicitly)

Writing Manager (triage + convergence)

Mission: make the draft converge by fixing only failing units and routing issues upstream when needed.

Do:

• Rewrite only files flagged by the gate (one at a time).
• If evidence is thin, route upstream instead of padding prose.
• Keep citation keys and scope stable during rewrites.

Avoid:

• Global rewrites that churn working sections.
• Rewriting around evidence gaps with generic filler.

Router (earliest responsible artifact)

Mission: connect a visible writing failure to the earliest upstream fix.

Do:

• Template voice -> local rewrite (subsection-polisher / front-matter-writer / chapter-lead-writer).
• Thin evidence -> evidence-selfloop (notes/bindings/packs).

Avoid:

• Repeatedly rewriting the same section when the pack cannot support it.

Role prompt: Writing Manager (triage + convergence)

You are the writing manager for a survey draft.

Your job is to make the draft converge by fixing only what fails, and by routing issues to the earliest responsible artifact.

Rules:
- If a section fails because the evidence is thin or out-of-scope pressure is high, do NOT pad prose.
  Route upstream to evidence-selfloop (notes/bindings/packs) and unblock the substrate.
- If a section fails because the prose is templated (narration openers, slide navigation, repeated disclaimer spam), rewrite locally.
- Never change citation keys during rewrites; keep scope local.

Working style:
- small, auditable edits
- one failing file at a time
- every fix either (a) improves argument moves, or (b) restores paper voice

• the gate is deterministic (it inspects sections/ + sections/sections_manifest.jsonl)
• the fixes are semantic (rewrite prose to execute argument moves and restore paper voice)

Inputs

• sections/sections_manifest.jsonl (expected files + allowed citations)
• sections/*.md (the actual prose units)
• outline/writer_context_packs.jsonl (preferred per-H3 pack)
• outline/evidence_bindings.jsonl
• citations/ref.bib

Optional routing context:

• output/EVIDENCE_SELFLOOP_TODO.md (if evidence gaps are already known)
• outline/subsection_briefs.jsonl, outline/chapter_briefs.jsonl

Output

• output/WRITER_SELFLOOP_TODO.md (report-class; always written)

Workflow

1) Run the gate

• Run the writer self-loop script (see Script section below). It reads sections/sections_manifest.jsonl and the actual sections/*.md files.

2) Read the TODO report

• Open output/WRITER_SELFLOOP_TODO.md and identify the failing file paths.
• Treat repeated evidence-sized sentences within one H3 as blocking; repair support selection or remove the duplicate without changing citation scope.
• For each failing H3, load its pack from outline/writer_context_packs.jsonl.
• If a pack is missing or you need scope reminders, consult outline/subsection_briefs.jsonl; for chapter-level context (leads/throughlines), consult outline/chapter_briefs.jsonl.

3) Triage (writing vs evidence)

Route upstream (run evidence-selfloop) when:

• output/EVIDENCE_SELFLOOP_TODO.md already indicates missing anchors/comparisons for the same subsection
• the pack in outline/writer_context_packs.jsonl is too thin to satisfy argument moves without guessing
• you keep wanting out-of-scope citations outside outline/evidence_bindings.jsonl

Rewrite locally when:

• the pack is rich enough, but prose is template-y (narration openers, slide navigation, repeated discourse stems)
• the section has content but lacks a thesis, contrasts, evaluation anchoring, or limitations

4) Use the right playbook for the failing file type

• Front matter (sections/abstract.md, sections/S<sec_id>.md for Intro/Related Work, sections/discussion.md, sections/conclusion.md): use front-matter-writer.
• Chapter leads (sections/S<sec_id>_lead.md): use chapter-lead-writer.
• H3 bodies (sections/S<sub_id>.md): use subsection-writer (draft) or subsection-polisher (local fix).

5) Keep scope + citations stable

• Keep citation keys unchanged and validate them against citations/ref.bib.
• Stay in scope per outline/evidence_bindings.jsonl (prefer subsection-first).

6) Rerun the gate until PASS

After PASS (merge-aware voice safety):

• Proceed to style-harmonizer -> opener-variator ->

section-logic-polisher -> paragraph-curator ->

evaluation-anchor-checker -> final argument-selfloop snapshot -> optional

transition-weaver -> section-merger -> post-merge-voice-gate.

• If post-merge-voice-gate FAILs with source: transitions, transition

insertion was explicitly enabled; fix outline/transitions.md and re-merge.

After PASS (mandatory style hygiene for all survey-family profiles):

• Open output/WRITER_SELFLOOP_TODO.md and read ## Style Smells.
• Run style-harmonizer on the listed sections/*.md files.
• If it flags opener cadence / overview narration, then run

opener-variator on the listed files.

• If it flags count-based limitation slots (e.g., Two limitations ...), run limitation-weaver on the listed files.
• After logic and paragraph compaction, run evaluation-anchor-checker on the

settled H3 files, then regenerate the argument/manifest snapshot.

• Keep meaning + citations fixed; treat this as surface-level rewrite, not new content.

How to fix a failing H3 (semantic recipe)

• Use tension_statement + thesis from the pack to rewrite paragraph 1 (end with the thesis).
• Use comparison_cards to write explicit A-vs-B contrast sentences.
• Use evaluation_anchor_minimal / protocol snippets to add task/metric/constraint context.
• Use limitation hooks to add a real caveat (not boilerplate).

Paper voice repairs (high-impact, low-risk)

Narration opener -> content claim:

• Bad: This subsection surveys ...
• Better: start with a tension/decision/lens sentence, then end paragraph 1 with the thesis.

Slide navigation -> argument bridge:

• Bad: Next, we move from X to Y.
• Better: Having established X, we can now examine how Y changes the trade-offs under comparable protocols.

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

• Bad: Therefore, survey comparisons should ...
• Better: Across reported protocols, ... varies, which makes ... fragile unless ...

Disclaimer spam -> one policy paragraph + local caveat only when needed:

• Keep evidence policy once in front matter; delete repeated \"abstract-only\" boilerplate inside H3.

Stop conditions (when rewriting is the wrong move)

Stop and route upstream if:

• you cannot write a contrast or evaluation anchor without guessing
• the pack lacks benchmarks/metrics/protocol details needed for the subsection\x27s core claim
• fixing one H3 keeps forcing out-of-scope citations

Failure codes -> routing (how to read the gate)

The writer gate emits short issue codes. Treat them as routers and stop trying to pad prose around missing evidence.

First classify failures:

• Evidence substrate: packs/anchors/comparisons are too thin to execute argument moves without guessing. Route upstream (evidence-selfloop).
• Writing execution: packs are rich enough, but the section did not execute moves (thesis/contrast/eval/limitation). Rewrite locally.
• Voice contamination: narration/slide/pipeline voice. Rewrite locally (opener + bridges), do not touch citations.
• Citation hygiene: undefined keys / out-of-scope pressure. Fix citations/bindings or rewrite to in-scope; do not add new keys.

Quick map (common codes):

• missing_sections_manifest, empty_sections_manifest, sections_missing_files: run subsection-writer (create missing files + refresh manifest).
• sections_h3_has_headings: remove headings; H3 bodies are body-only.
• sections_intro_*, sections_related_work_*: rewrite front matter via front-matter-writer (dense positioning + one methodology note).
• sections_h3_too_few_paragraphs, sections_h3_too_short: if the pack is rich, expand by executing contrasts + eval anchor + limitation; if the pack is thin, route to evidence-selfloop.
• sections_h3_missing_contrast: add explicit A-vs-B using comparison_cards; if cards are missing, route to evidence-selfloop (briefs/packs).
• sections_h3_missing_eval_anchor, sections_h3_missing_cited_numeric: add minimal protocol context (task/metric/constraint) in the same paragraph; if unknown, route upstream or weaken the claim.
• sections_h3_missing_limitation: add a subsection-specific limitation; if none exists in the pack, route upstream.
• sections_h3_narration_template_opener, sections_h3_slide_narration, sections_contains_pipeline_voice: rewrite openers/bridges to paper voice (no navigation commentary).
• sections_h3_evidence_policy_disclaimer_spam: delete repeats; keep evidence policy once in front matter.
• sections_cites_missing_in_bib: rerun citation-verifier (do not invent keys).
• sections_cites_outside_mapping: rewrite to in-scope OR fix mapping/bindings (C2/C4) and rebuild packs; do not patch in the merged draft.
• sections_h3_sparse_citations: if scope allows, plan adds via citation-diversifier then apply via citation-injector; if scope is too tight, expand mapping/bindings upstream.

Script (optional; deterministic gate)

Quick Start

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

All Options

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

Examples

• Generate an actionable TODO list for failing sections/*.md:
• uv run python .codex/skills/writer-selfloop/scripts/run.py --workspace <workspace>