Writer Selfloop

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass, no license.

.claude/skills

Audit 0.4.1 · source sha 81994516a430 · 19 Oct 2026Report an issue or request a re-scan

• PublisherWILLOSCAR GitHub userOwner of the repository this skill is read from
• Source folderWILLOSCAR/research-units-pipeline-skills/.codex/skills/writer-selfloopThe exact folder we audited
• LicenceNot declaredAs the repository states it
• Commit pinned for this verdict81994516a430A new audit runs when the source changes
• Repository statusNot archivedLast push 14 days ago
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Writing self-loop for surveys: run the strict section-quality gate, then rewrite only the failing `sections/*.md` files until the report is PASS.

---
name: writer-selfloop
description: "Writing self-loop for surveys: run the strict section-quality gate, then rewrite only the failing `sections/*.md` files until the report is PASS."
---

# Writer Self-loop (fix only what fails)

## Triggers & routing

- **Trigger**: writer self-loop, writing loop, quality gate loop, rewrite failing sections, 自循环, 反复改到 PASS.
- **Use when**: per-section files exist but C5 is FAIL/BLOCKED (thin sections, missing leads/front matter, citation-scope violations, generator voice).


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)

```text
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 `## 

Excerpt, first 6,000 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

1. 2026-10-19
   ✓SAFE B 89 / 100

   first audit · source 81994516a430

Questions about Writer Selfloop

• github.com/WILLOSCAR/research-units-pipeline-skills/tree/main/.c — the audited source at commit 81994516a430 and every finding aboveObserved · 19 Oct 2026

Spotted something wrong? Report a correction · Methodology