本文へ移動
cccskills
無料GitHub で公開

outline-budgeter

Merge/simplify an over-fragmented outline to hit a paper-like section budget (NO PROSE): target final ToC ~6–8 H2, fewer thicker H3.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md3.4 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Outline Budgeter (NO PROSE)

Triggers & routing

  • Trigger: outline budget, merge sections, too many sections, H3 explosion, 大纲预算, 合并小节, 大纲太碎.
  • Use when: outline/outline.yml exists but would produce thin sections (too many H2/H3); before (or immediately after) section-mapper.

Purpose: prevent the most common survey failure mode: H3 explosion (too many tiny subsections) leading to a thin, outline-like PDF.

This skill rewrites outline/outline.yml into a paper-like budget:

  • Final ToC target: ~6–8 H2 sections (Intro / Related Work + 3–4 core chapters + Discussion + Conclusion)
  • H3 target: fewer, thicker subsections (per draft_profile: course_paper<=6, survey<=10, deep<=12)

Important: Discussion/Conclusion are appended in C5 merge (global sections), so the outline itself should usually be <=6 H2.

Inputs

  • outline/outline.yml
  • Optional (helps make merges evidence-aware):
    • queries.md (optional: if it sets draft_profile, use it to choose the H3 budget)
    • outline/mapping.tsv
    • outline/coverage_report.md
    • GOAL.md

Outputs

  • outline/outline.yml (updated in place)
  • outline/OUTLINE_BUDGET_REPORT.md (bullets-only; what was merged and why)

Workflow (NO PROSE)

  1. Read the outline and compute a simple budget snapshot:
  • If queries.md sets draft_profile (course_paper/survey/deep), use it to decide the H3 budget target.
  • H2 count (excluding Discussion/Conclusion, which are not in the outline)
  • Total H3 count
  • H3 count per H2 chapter
  1. Decide a merge plan (structure-first, evidence-aware):
  • Prefer merging adjacent H3s that share similar axes/keywords.
  • Prefer merging H3s with weak mapping coverage (if mapping.tsv exists).
  • If outline/coverage_report.md exists (from outline-refiner), use it to identify weak-coverage or high-reuse subsections to merge.
  • Use GOAL.md as the scope constraint: avoid merges that mix distinct research questions or scope boundaries.
  • Prefer moving fine-grained distinctions into bullets/axes instead of creating new subsections.
  1. Apply merges in outline/outline.yml:
  • Merge titles into a clearer, thicker subsection title.
  • Merge bullets (dedupe templates; keep Stage A fields: Intent/RQ/Evidence needs/Expected cites).
  • Keep ids stable when possible.
    • If you must change ids, record it explicitly in the report and assume mapping.tsv must be regenerated.
  1. Write outline/OUTLINE_BUDGET_REPORT.md:
  • Before/after counts.
  • List of merges (old ids/titles -> new id/title).
  • Any risks (e.g., mapping reset required).

Quality checklist

  • No placeholders (TODO/…/(placeholder)).
  • Outline budget matches the paper-like target.
  • Each remaining H3 is thick enough to sustain evidence-first writing (its bullets mention concrete comparisons + eval anchors + failure modes).

Troubleshooting

Issue: merging makes a subsection too broad

Fix:

  • Keep one H3, but split its bullets into explicit comparison axes and required evidence fields; defer fine-grained splits to later if evidence is strong enough.

Issue: mapping breaks after id changes

Fix:

  • Rerun section-mapper to regenerate outline/mapping.tsv, then rerun outline-refiner.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Internal template for creating or refactoring a skill into the repository's reference-first shape. **Trigger**: reference-first template, blueprint skill, create a reusable skill, refactor a script-heavy skill. **Use when**: you need a lean `SKILL.md`, explicit `references/`, machine-readable `assets/`, and a minimal deterministic `run.py`.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

Download a small corpus of open-access arXiv survey/review PDFs about agentic systems and extract text for style learning.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

Extract per-subsection “anchor facts” (NO PROSE) from evidence packs so the writer is forced to include concrete numbers/benchmarks/limitations instead of generic summaries.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

Curate reader-facing survey tables for the Appendix (clean layout + high information density), using only in-scope evidence and existing citation keys.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

Build a final, machine-readable argument snapshot from current H3 sections and refresh section fingerprints before merge.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

Audit one research Workspace for declared Unit outputs and Pipeline target Artifacts, writing `output/CONTRACT_REPORT.md`; use for mid-Run coverage snapshots or final delivery completeness, not deep provenance integrity.

日本語の概要は準備中です。原文の説明を表示しています。

WILLOSCAR/research-units-pipeline-skills5132026年10月5日 更新

WILLOSCAR のスキルをすべて見る

このスキルの問題を報告する