Writer Context Pack

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

digest = hashlib.sha1((key or "").encode("utf-8", errors="ignore")).hexdigest()

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-context-packThe 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: Build per-H3 writer context packs (NO PROSE): merge briefs + evidence packs + anchor facts + allowed citations into a single deterministic JSONL, so drafting is less hollow and less brittle.

---
name: writer-context-pack
description: "Build per-H3 writer context packs (NO PROSE): merge briefs + evidence packs + anchor facts + allowed citations into a single deterministic JSONL, so drafting is less hollow and less brittle."
---

# Writer Context Pack (C4→C5 bridge) [NO PROSE]

## Triggers & routing

- **Trigger**: writer context pack, context pack, drafting pack, paragraph plan pack, 写作上下文包.


Purpose: reduce C5 “hollow writing” by giving the writer a **single, per-subsection context pack**:
- the exact RQ/axes + paragraph plan (`subsection_briefs`)
- concrete comparison cards + evaluation protocol + limitations (`evidence_drafts`)
- numeric/eval/limitation anchors (`anchor_sheet`)
- allowed citation scope (subsection + chapter union) from `evidence_bindings`

## Inputs

- `outline/outline.yml`
- `outline/subsection_briefs.jsonl`
- `outline/chapter_briefs.jsonl`
- `outline/evidence_drafts.jsonl`
- `outline/anchor_sheet.jsonl`
- `outline/evidence_bindings.jsonl`
- `citations/ref.bib`

## Read as needed

- `references/source_text_hygiene.md` when bad source sentences or paper self-narration leak into writer packs

## Machine-readable assets

- `assets/paper_voice_palette.json`
- `assets/source_text_hygiene.json`
- `assets/context_pack_policy.json`
- repo-wide `assets/limitation-signals.json` — shared polarity rules used before
  a sentence can become a limitation hook

## Outputs

- `outline/writer_context_packs.jsonl`

## Output format (`outline/writer_context_packs.jsonl`)

JSONL, one object per H3 subsection.

Required keys:
- `sub_id`, `title`, `section_id`, `section_title`
- `rq`, `thesis`, `axes`, `paragraph_plan`
- `clusters` (copied from subsection briefs so writer-side evidence can still be assigned per route when comparison cards are sparse)
- `tension_statement`, `evaluation_anchor_minimal` (copied from subsection briefs; concrete tension + minimal eval context slots)
- `opener_mode`, `opener_hint` (paper-voice hint to vary subsection openers without template labels)
- `bridge_terms`, `contrast_hook`, `required_evidence_fields` (copied from subsection briefs; transition/evidence handles; NO NEW FACTS)
- `chapter_synthesis_mode` (copied from chapter briefs; helps avoid template-y “Taken together...” repeats)
- `allowed_bibkeys_{selected,mapped,chapter,global}`
- `anchor_facts` (trimmed)
- `comparison_cards` (trimmed)
- `must_use` (writer contract; minima derived from pack richness + `draft_profile`)
- `do_not_repeat_phrases` (rewrite triggers; high-signal generator stems to avoid repeating verbatim)
- `pack_warnings` (list; why this pack may still draft hollow if not fixed upstream)
- `paper_voice_palette` (positive paper-voice phrase palette + rewrite stems; avoids "generator voice" without brittle hard blocks)
  - Source: workspace override `outline/paper_voice_palette.json` (if present), else repo default `.codex/skills/writer-context-pack/assets/paper_voice_palette.json`.
  - Includes `role_cards` (section author / evidence steward / style harmonizer). Treat them as role cues and rewrite intentions, not sentence templates.
- `pack_stats` (object; raw/kept/dropped counts + trim policy so truncation/drop is not silent)

Trim policy:
- Packs trim long snippets to preserve concrete details while keeping JSONL readable (see `pack_stats.trim_policy`).
- Trimming does not add ellipsis markers (to reduce accidental leakage into prose).

## Writer contract (how C5 should use this pack)

Treat each pack as an executable checklist, not optional context:

A150++ minima (defaults; used by gates and self-loops):
- paragraph_plan: 10
- anchor_facts: >=10
- comparison_cards: >=4
- limitation_hooks: >=3
- allowed_bibkeys_mapped: >=28

`course_paper` minima use a smaller context pack: paragraph_plan 6,
anchor_facts >=6, comparison_cards >=3, limitation_hooks >=2, and
allowed_bibkeys_mapped >=8. The writer-facing `must_use` subset is 3 anchors,
2 comparisons, and 1 limitation.

- **Plan compliance**: follow `paragraph_plan` (don’t skip planned paragraphs; merge only if you keep the same contrasts/anchors).
- **Connector intent**: treat `paragraph_plan[].connector_phrase` as semantic guidance, not copy-paste; paraphrase and vary; avoid `Next, we ...` narration.
- **Anchors are must-use**: include at least one `anchor_facts` item that matches your paragraph’s claim type (eval / numeric / limitation), when present.
- **Comparisons are must-use**: reuse `comparison_cards` to write explicit A-vs-B contrast sentences (avoid “A then B” separate summaries).
- comparison cards are only valid when both sides carry usable highlights; if a card becomes one-sided after hygiene, drop it and route the gap upstream instead of asking the writer to improvise the missing side
- **Thesis is must-use**: the first paragraph should end with the `thesis` statement (or a faithful paraphrase with the same commitment level).
  - Prefer a content claim; avoid generator-like meta openers (`This subsection ...`) and avoid repeating literal opener labels (e.g., `Key takeaway:`) across many H3s.
- **Opener mode (anti-template)**: use `opener_mode` / `opener_hint` to vary how paragraph 1 frames the subsection (tension-first vs decision-first vs lens-first).
  - Do not copy labels into the prose; keep signposting light and content-bearing.
- **Anti-template**: treat `do_not_repeat_phrases` as rewrite triggers (paper voice hygiene):
  - if a listed phrase appears, rewrite it into a content claim / argument bridge (no outline narration)
  - avoid swapping in a new repeated stem; keep phrasing varied and paper-like
  - exception: a phrase may appear once if it is truly subsection-specific (not a reusable stem)
- **Micro-structure**: if prose starts drifting into flat summaries, apply `grad-paragraph` repeatedly (tension → contrast → evaluation anchor → limitation).
- **Citation scope**: prefer `allowed_bibkeys_selected` (then `allowed_bibkeys_mapped`, then `allowed_bibkeys_chapter`). `allowed_bibkeys_global` is reserve

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 Context Pack

• 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