本文へ移動
cccskills
無料GitHub で公開

paper-formal-writer

Plan, draft, audit, globally revise, format, and verify a formal mathematical-modeling paper after the Standard evidence gate passes. Use for the sole S7/S8 formal manuscript path, not for legacy micro-unit scaffolds.

インストール方法を見る

含まれるファイル(16)

  • SKILL.md6.4 KB
  • agents/openai.yaml335 B
  • references/competition-scope.md2.1 KB
  • references/cumcm-paper-standard.md3.3 KB
  • references/figure-table-writing-rules.md1.2 KB
  • references/formal-paper-template.md1.7 KB
  • references/section-expansion-rules.md1.3 KB
  • scripts/assemble_sections.py3.2 KB
  • scripts/authoring_contracts.py15.0 KB
  • scripts/build_paper_outline.py14.4 KB
  • scripts/check_paper_format.py36.1 KB
  • scripts/format_formal_docx.py28.6 KB
  • scripts/formula_omml.py9.6 KB
  • scripts/paper_scope.py4.3 KB
  • scripts/prepare_authoring.py6.8 KB
  • scripts/validate_authoring.py16.2 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。

Paper Formal Writer

This skill is the sole formal author and final manuscript producer in Standard. It turns a fresh S6 evidence chain into an audited Markdown source and native-equation Word document.

Entry

Run before formal work:

python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --skill paper-formal-writer

S0-S6 must pass. If workflow state is uncertain, run workflow_guard.py --status and follow its current artifact report rather than memory.

Single Formal Path

S6 evidence PASS
-> writing plan
-> chapter drafts
-> chapter audits
-> queued local repair when required
-> deterministic assembly
-> Agent global revision
-> final Markdown audit
-> formal DOCX
-> S8 format/render gate

Do not promote quickstart, legacy, or mechanically merged micro-unit output into a formal filename.

Prepare

  1. Build the evidence-aware outline:
python .claude/skills/paper-formal-writer/scripts/build_paper_outline.py
  1. Prepare adaptive authoring:
python .claude/skills/paper-formal-writer/scripts/prepare_authoring.py --mode auto

Default delivery is competition: plan at least 14000 effective characters, with a main-paper floor of 8000 and at least 18 rendered pages before appendices. These are adjustable project defaults, not contest rules. Verify the actual contest's page cap and counting rules before writing. auto therefore selects section; a short report needs --delivery short-report --scope-reason "user-requested scope". Custom floors use --min-pages / --min-body-chars with a reason; never lower them merely to turn a failed check green.

Read references/competition-scope.md when preparing scope or repairing a short manuscript. Several authoring turns may produce one coherent chapter; preserve evidence and notation between turns instead of compressing the paper into one response.

The authoritative S7 contracts are:

  • paper_output/plan/writing_plan.json
  • paper_output/context/authoring_state.json
  • paper_output/qa/draft_audit.json
  • paper_output/qa/repair_queue.json

Draft And Audit

Write only to the path recorded for the current unit in authoring_state.json. Each chapter must be a complete, coherent section and must include its required removable marker:

<!-- mathmodel-evidence: evidence-id-1, evidence-id-2 -->

Audit every changed draft:

python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --section <section-id>

Routing is deterministic:

  • global repeats the same blocking category twice: switch to section.
  • A section fails once: rewrite that section.
  • A section repeats the same category twice: $paper-micro-unit-generator may repair only queued locations.
  • The third changed attempt with the same category: S7 becomes BLOCKED; report the cause and suggest Lite as a user choice without switching automatically.

Audits block short drafts, missing chapters/evidence/numbers, broken formulas, placeholders, duplicate prose, broken figure/table references, and internal workflow language.

Assemble And Revise

After all active units pass:

python .claude/skills/paper-formal-writer/scripts/assemble_sections.py
python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --assembled

The assembler follows outline order, strips evidence comments, and preserves existing numbering. It writes paper_output/drafts/assembled_draft.md.

Then read the entire assembled draft and evidence chain, and rewrite the paper globally into:

paper_output/final_paper_source.md

This pass must unify terminology, notation, transitions, argument order, citations, captions, and conclusions. Copying the assembled draft unchanged is rejected.

Run the final source audit:

python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --final

Any change to evidence, writing plan, approved section draft, assembly, or final source invalidates downstream PASS state.

Word And S8

Only after authoring_state.status = PASS:

python .claude/skills/paper-formal-writer/scripts/format_formal_docx.py
python .claude/skills/paper-formal-writer/scripts/check_paper_format.py --render required

Formal mode requires fresh S6 and S7 hashes. --allow-draft may create final_paper_draft.docx for layout diagnosis, but never satisfies S7/S8 or overwrites the formal DOCX.

format_formal_docx.py converts LaTeX to editable Word OMML and blocks failed conversion. check_paper_format.py checks dynamic length, required hierarchy, citations, evidence-linked figures/tables, formulas, duplicate prose, DOCX structure, LibreOffice PDF rendering, page count, and extractable text.

Writing Standard

  • Use 1 / 1.1 / 1.1.1 headings and complete CUMCM sections.
  • For each question, connect assumptions, variables, derivation, algorithm, computed values, validation, uncertainty, and conclusion.
  • Define symbols before formulas and explain every displayed result.
  • Cite and interpret every included figure/table.
  • Use only current evidence; never invent runs, values, sources, or validation.
  • Avoid copied or number-swapped paragraphs used only to meet length.
  • Keep implementation language out of body prose; reproducibility paths belong in the appendix.

References

Read only the reference needed for the current writing or repair decision.

Handoff

After each meaningful transition, run workflow status and update paper_output/context/workflow_memory.json:

python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --status
python .claude/skills/context-memory-keeper/scripts/update_workflow_memory.py

S8 is complete only when the machine-readable format report is PASS and fresh.

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

自动定位并获取权威公开数据(优先API/官方批量下载),输出可复现抓取与清洗方案。Invoke when用户需要权威数据、官方统计、API下载或数据源爬取。

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

Manages persistent memory. Invoke to read active context or archive old tasks. Structure: Long-term Principles (Rules) + Short-term Workbench (Tasks).

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

自动清洗赛题或爬取的数据(处理缺失/异常/格式),并生成可视化图表。Invoke when 用户需要处理原始数据、清洗数据或生成数据分析图表。

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

根据 model_route.json、数据计划和清洗数据,为数学建模论文生成结果证据契约和 q1/q2/q3 建模代码脚手架。Invoke when 需要把模型输出、评价指标、结构化结论、论文表格和当前赛题专用建模代码沉淀到 paper_output/results/、paper_output/tables/ 和 paper_output/code/modeling/,供 QA 与正文生成读取。

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

按常见评分点生成建模论文结构与写作清单,并根据题目类型与数据条件给出模型选择与对照实验路线。Invoke when需要“论文格式/评分对齐/模型选型/路线不确定”。

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

Repair a repeatedly failing Standard paper section from repair_queue.json, or generate an explicitly requested legacy/quickstart scaffold. Do not use it as the formal paper writer.

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。

yushui2022/MathModel-Skill4542026年10月7日 更新

yushui2022 のスキルをすべて見る

このスキルの問題を報告する