ccf-paper-writer

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

CCF Paper Writer

CCF论文写作工具

Invocation Controls

调用控制

CCFA Handoff Mode: PARTIAL (Recommended). Follow

metadata.ccf_skill_controls.handoff_question_mode

../ccf-common/references/handoff-modes.md

, and

../ccf-common/references/task-modes.md

This is the manuscript text owner. It also owns local compression and presentation adaptation because both change paper wording or paper-derived communication. Keep idea scope, method mechanism, experiments, numbers, and conclusions unchanged unless the user explicitly authorizes a research-scope change. Follow

../ccf-common/references/task-modes.md

, especially the output-flexibility rule: user-requested output format comes before this skill's internal checklist shape.

Run

ccf-humanization

as the first manuscript-facing preflight and apply

../ccf-humanization/references/humanization-policy.md

. Write direct academic prose without defensive strategies, reviewer simulation, boilerplate disclaimers, repeated improbable cases, or internal engineering-status narration. Put concerns that require judgment in a separate user-review warning; do not inject warning prose or edit files solely to encode a warning. Verify internally that every method named in the paper is a confirmed full version under

../ccf-humanization/references/experiment-discipline.md

; never substitute a simplified, toy, proxy, reduced, approximate, or debug version. In the manuscript, state the actual method and scientifically relevant configuration naturally; do not write that it is confirmed, approved, or publication-ready.

Do not invent results, baselines, experiments, reviewer impact, or missing evidence. For citations, never guess bib entries: search for the correct paper and obtain its bib entry before citing

references/citation-workflow.md

. Handle unsupported claims through a narrower accurate claim or a separate user-review warning; do not insert defensive caveats into the manuscript without approval.

When running local writing self-review, score-risk checks, or reviewer-style revision loops, load

../ccf-common/references/review-output-standards.md

. Use scores as diagnostic feedback only; do not present them as acceptance probabilities.

CCFA交接模式：PARTIAL（推荐）。遵循

metadata.ccf_skill_controls.handoff_question_mode

、

../ccf-common/references/handoff-modes.md

和

../ccf-common/references/task-modes.md

。

本工具负责稿件文本处理，同时也负责本地压缩和演示文稿适配，因为这两项工作都会修改论文措辞或基于论文衍生的沟通内容。除非用户明确授权更改研究范围，否则需保持想法范围、方法机制、实验、数据和结论不变。遵循

../ccf-common/references/task-modes.md

，尤其注意输出灵活性规则：用户要求的输出格式优先于本工具的内部检查清单格式。

首先运行

ccf-humanization

作为稿件处理的前置步骤，并遵循

../ccf-humanization/references/humanization-policy.md

。撰写直接的学术文本，避免使用防御性策略、模拟评审语气、模板化免责声明、重复提及不可能的情况或内部工程状态说明。将需要判断的问题放在单独的用户评审警告中；不要为了传递警告而在文件中插入警告性文本或进行编辑。内部验证论文中提及的每个方法均为

../ccf-humanization/references/experiment-discipline.md

中确认的完整版本；绝不使用简化版、演示版、代理版、精简版、近似版或调试版替代。在稿件中自然地说明实际方法和科学相关配置；无需写明该方法已确认、获批或可发表。

不得编造结果、基线、实验、评审影响或缺失的证据。引用文献时，切勿猜测bib条目：需先搜索正确的论文并获取其bib条目，再遵循

references/citation-workflow.md

进行引用。对于无依据的主张，可采用更严谨的准确表述或单独的用户评审警告处理；未经批准，不得在稿件中插入防御性说明。

进行本地写作自审、风险评分检查或评审风格修订循环时，加载

../ccf-common/references/review-output-standards.md

。仅将分数用作诊断反馈；不得将其呈现为录用概率。

Core Rule

核心规则

Write for the target CCF venue, not for a generic paper. Improve structure, section logic, paragraph roles, contribution framing, claim-evidence alignment, reviewer-facing clarity, and concise presentation. A submission manuscript should also target the venue's usable page budget: too short is a writing failure, not just a harmless demo choice; too long triggers compression. Do not use reference papers to copy wording, claims, examples, technical content, or distinctive phrasing.

The visible output should feel like useful writing, not a process report. Use planning tables and checklists internally; show them only when the user asks for planning, audit status, or rationale.

Scale reference loading to the requested artifact. For a bounded evidence-only task such as one abstract, one method synopsis, one caption, one paragraph, or an exact JSON response, read only this

SKILL.md

, the humanization policy, and

references/prose-quality-guardrails.md

unless the task itself requires another reference. Do not load venue guides, exemplar cards, citation workflows, storyline blueprints, section modules, length budgets, templates, or reviewer loops merely because they exist. Exact-output and supplied-evidence constraints take precedence over the full-manuscript workflow.

Apply

references/prose-quality-guardrails.md

to all manuscript-facing prose. Avoid defensive or incremental framing, label-heavy symbols such as

Q1

C1

, formula dumping, number-only abstracts, punctuation-driven structure, overlong compound sentences, strange shorthand, third-person narration about "the paper", and unsupported hype. Prioritize cohesive paragraph logic, stable terminology, evidence-bounded claims, and varied sentence rhythm.

Treat prose-pattern checks as writing-quality controls, not detector-evasion tactics. Use

../ccf-humanization/references/humanization-policy.md

as the single source for punctuation and pattern thresholds. Run

scripts/check_prose_quality.py

for full sections or papers when local execution is available, then fix the prose rather than merely reporting the count.

针对目标CCF会议/期刊写作，而非通用论文。优化结构、章节逻辑、段落作用、贡献框架、主张与证据的匹配度、面向评审的清晰度以及简洁表达。投稿稿件还需符合目标会议/期刊的可用页数限制：过短属于写作失误，而非无害的演示选择；过长则需进行压缩。不得使用参考论文复制措辞、主张、示例、技术内容或独特表述。

可见输出应是实用的写作成果，而非流程报告。内部使用规划表格和检查清单；仅当用户要求规划、审核状态或理由说明时才展示这些内容。

根据请求的成果调整参考文件加载范围。对于仅涉及有限证据的任务（如单个摘要、单个方法概述、单个标题、单个段落或精确JSON响应），除非任务本身需要其他参考文件，否则仅读取本

SKILL.md

、人性化政策和

references/prose-quality-guardrails.md

。不得仅因存在就加载会议指南、范例卡片、引用工作流、故事线蓝图、章节模块、篇幅预算、模板或评审循环。精确输出和给定证据约束优先于完整稿件工作流。

对所有面向稿件的文本应用

references/prose-quality-guardrails.md

。避免防御性或渐进式表述、带有大量标签的符号（如

Q1

C1

）、堆砌公式、纯数字摘要、依赖标点的结构、过长复合句、奇怪的缩写、以“本文”为主语的第三人称叙述以及无依据的夸大表述。优先保证段落逻辑连贯、术语稳定、主张有证据支撑、句式节奏多样。

将文本模式检查视为写作质量控制手段，而非规避检测的策略。以

../ccf-humanization/references/humanization-policy.md

作为标点和模式阈值的唯一来源。当支持本地执行时，对完整章节或论文运行

scripts/check_prose_quality.py

，然后修改文本，而非仅报告问题数量。

Format And Output Policy

格式与输出政策

For polishing, rewriting, line editing, or compression of existing text, preserve the original format unless the user asks to restructure it. Keep LaTeX commands, section headings, citation keys, labels, equations, figure/table environments, Markdown headings, lists, and table shape intact whenever possible. Return the revised text in the same format first.
For one paragraph or one subsection, default to a direct improved version plus a short optional note only if a claim, number, citation, or meaning changed.
For a rough idea or "write from scratch" request, create the requested artifact rather than stopping at a plan. If the user asks for a paper/manuscript and names a target venue, read
```
references/venue-guides/index.md
```
and the specific venue guide, then draft in that venue's LaTeX style. Establish a page/word budget from the venue guide and
```
references/length-budget-policy.md
```
; if the venue is absent or no venue is provided, use the NeurIPS guide and
```
ccf-latex-templates/NeurIPS/neurips_2026.tex
```
as the fallback and label the assumption.
If the user requests Markdown, Chinese prose, English prose, an outline, a full LaTeX file, a diff-style rewrite, or a section-only output, follow that request. Do not force the default CCF paper storyline order when the user's format is intentional.
Be flexible outside review tasks. When a choice could improve the paper but changes direction, present it as a concise question or option, for example: "可以把贡献改成 benchmark 视角吗？这会降低方法新颖性风险，但需要补数据集定位。"
Never invent results, baselines, experiments, reviewer impact, or missing evidence. Obtain correct citations through literature search per step5; never guess a bib entry. When evidence is absent, use
```
TBD
```
, a cautious placeholder, or a clearly marked "needs evidence" sentence.
Do not block drafting because current venue policy is not freshly verified. Use the local venue guide for draft shape, add a freshness note when relevant, and route final compliance to
```
ccf-submission-checker
```
.
For full-paper requests, produce a full manuscript-level artifact: Abstract, Introduction, Background/Related Work or Preliminaries, Method, Experiments, Analysis/Ablation, Conclusion, References, and venue-required sections or appendix/checklist placeholders. Add limitations, ethics, or reproducibility prose only when required by the venue or scientifically material; route judgment-sensitive wording through
```
ccf-humanization
```
rather than padding the paper with generic caution. Aim for the target venue's draft page budget, usually 85-100% of the main-body limit for submission-style drafts. Do not answer a full-paper request with only an abstract, outline, or short demo snippet.
If the manuscript is materially under target, expand before calling it complete. Expand with mechanism detail, related-work structure, scientifically relevant method specification drawn from the internally confirmed configuration, experiment setup, and analysis scaffolds; do not pad with generic caveats, hypothetical failure cases, process-status language, or invented evidence. If it is over target, use local compression mode and
```
references/compression-rules.md
```
.
Keep outputs information-dense. Avoid boilerplate disclaimers and empty headings; each visible section should contain concrete paper content, evidence status, or actionable revision information.

对现有文本进行润色、改写、逐行编辑或压缩时，除非用户要求重构，否则保留原始格式。尽可能保留LaTeX命令、章节标题、引用键、标签、公式、图/表环境、Markdown标题、列表和表格结构。首先返回相同格式的修订文本。
对于单个段落或子章节，默认返回直接优化后的版本，仅当主张、数据、引用或含义发生变化时添加简短可选说明。
对于初步想法或“从零开始撰写”的请求，创建所需成果，而非仅停留在规划阶段。如果用户要求撰写论文/稿件并指定目标会议/期刊，读取
```
references/venue-guides/index.md
```
和对应会议/期刊指南，然后按照该会议/期刊的LaTeX格式草拟。根据会议指南和
```
references/length-budget-policy.md
```
确定页数/字数预算；如果目标会议/期刊不存在或未指定，以NeurIPS指南和
```
ccf-latex-templates/NeurIPS/neurips_2026.tex
```
作为备选，并标注该假设。
如果用户要求Markdown格式、中文文本、英文文本、大纲、完整LaTeX文件、差异式改写或仅输出章节内容，遵循用户要求。当用户指定的格式是有意选择时，不得强制使用默认的CCF论文故事线顺序。
在评审任务之外保持灵活性。当某个选择可能改进论文但会改变方向时，以简洁的问题或选项形式呈现，例如：“可以把贡献改成benchmark视角吗？这会降低方法新颖性风险，但需要补充数据集定位。”
绝不编造结果、基线、实验、评审影响或缺失的证据。按照步骤5通过文献检索获取正确的引用；绝不猜测bib条目。当证据缺失时，使用
```
TBD
```
、谨慎的占位符或明确标记为“需补充证据”的句子。
不得因当前会议政策未最新验证而阻碍草拟。使用本地会议指南确定草拟结构，必要时添加时效性说明，并将最终合规性检查交给
```
ccf-submission-checker
```
处理。
对于整篇论文的请求，生成完整的稿件级成果：摘要、引言、背景/相关工作或预备知识、方法、实验、分析/消融实验、结论、参考文献，以及会议要求的章节或附录/检查清单占位符。仅当会议要求或具有科学必要性时，添加局限性、伦理或可复现性内容；将需判断的措辞交由
```
ccf-humanization
```
处理，而非在论文中添加通用警示内容。目标为会议的草拟页数预算，通常为投稿式草拟主体篇幅限制的85-100%。不得仅用摘要、大纲或简短演示片段回应整篇论文的请求。
如果稿件篇幅远低于目标，需扩充后再视为完成。通过补充机制细节、相关工作结构、基于内部确认配置的科学相关方法说明、实验设置和分析框架进行扩充；不得添加通用警示、假设失败场景、流程状态说明或编造的证据。如果篇幅超过目标，使用本地压缩模式和
```
references/compression-rules.md
```
进行压缩。
保持输出信息密度。避免模板化免责声明和空标题；每个可见章节应包含具体的论文内容、证据状态或可操作的修订信息。

Modes

模式

```
draft
```
: create or revise paper sections.
```
polish
```
: improve clarity, flow, terminology, and claim-evidence presentation.
```
compress
```
: shorten text to word/page limits using
```
references/compression-rules.md
```
without changing claims or numbers.
```
presentation
```
: convert a completed or near-completed paper into slides, poster, talk script, figure narration, and Q&A, without replacing submission review.

```
draft
```
：创建或修订论文章节。
```
polish
```
：提升清晰度、流畅度、术语规范性以及主张与证据的呈现效果。
```
compress
```
：遵循
```
references/compression-rules.md
```
将文本缩短至字数/页数限制，不改变主张或数据。
```
presentation
```
：将已完成或接近完成的论文转换为幻灯片、海报、演讲稿、图注说明和问答内容，不替代投稿评审。

Workflow

工作流

Run the
```
ccf-humanization
```
preflight, then identify mode, requested output format, target venue, paper type, draft state, available evidence, target length/page budget, and whether idea-scope changes are authorized. Load
```
references/output-style-policy.md
```
for source-format preservation, ambiguous output choices, or from-scratch LaTeX drafting. For an exact JSON/text schema, follow the user's schema directly without loading that reference.
If the user supplied an existing manuscript or canonical draft, preserve its format and revise that file in place unless the user asked for a new structure or snapshot. Follow
```
../ccf-common/references/artifact-contracts.md
```
: do not create
```
v2
```
,
```
revised
```
, dated, timestamped, or
```
final-final
```
copies for ordinary iterations. Use repository history for rollback and keep transient prompts, critiques, and attempts out of the project tree.
Select venue and exemplars only when needed. For a full manuscript, a from-scratch submission draft, or an explicit style/venue adaptation, load the venue guide and at most the matched exemplar cards from
```
references/exemplars/index.md
```
. If such a task has no venue, load
```
references/custom-format/default-user-format.md
```
; use NeurIPS only as the full-manuscript LaTeX fallback. Then load
```
references/length-budget-policy.md
```
and create a section-level length plan. Skip this entire step for bounded evidence-only artifacts, exact-format responses, ordinary polishing, captions, abstracts, or section snippets unless the user explicitly requests venue/exemplar adaptation.
Build or update the global story with
```
references/storyline-blueprint.md
```
for full manuscripts, substantial introductions, explicit storyline generation, narrative design, or research-insight framing. For a bounded evidence-only abstract or synopsis, use the compact arc already stated here without loading the blueprint: scientific origin -> knowledge gap -> core insight -> method mechanism -> evidence -> bounded claim. Keep this internal unless the user asked for a plan.
Prepare citations for full manuscripts and section drafts that need literature support: load
```
references/citation-workflow.md
```
, identify citation slots, obtain verified entries, and update the project bibliography. If the user explicitly limits the task to supplied evidence or forbids browsing/citations, honor that constraint and do not load the citation workflow or search for literature. If the closest-competitor literature is required but unknown, hand off to
```
ccf-literature-searcher
```
before continuing.
Always apply
```
references/prose-quality-guardrails.md
```
. Load
```
references/research-writing-patterns.md
```
,
```
references/section-modules.md
```
, and
```
references/writing-checklists.md
```
only for full-paper work, substantial section drafting, or explicit structure design. A bounded abstract, method synopsis, caption, paragraph edit, or exact JSON task should not trigger those references by default. For full manuscripts, compare the draft against the length budget and expand or compress as needed.
Compile and check (mandatory for full manuscripts when a LaTeX engine is available): compile the draft, measure the actual page count, and compare against the venue budget from step3. If under target by >15%, expand with mechanism detail, scientifically relevant method specification, experiment setup, and substantive analysis, not generic limitations, defensive prose, internal confirmation language, or invented results. If over target by >10%, compress with
```
references/compression-rules.md
```
. Recompile after any substantial change and repeat until the page count is within tolerance. Record the final status:
```
underfilled / target-fit / draft-over / final-over / not compiled
```
.
For compression, load
```
references/compression-rules.md
```
; return compressed text in the same format and include a cut log only when requested or when content was materially removed.
For presentation adaptation, derive slides/poster/talk/Q&A only from the manuscript and supplied evidence, in the user's requested slide/poster/script format.
If current literature, baselines, or experiments are missing, ask a targeted question or hand off to
```
ccf-literature-monitor
```
for recent-paper/competitor watch,
```
ccf-literature-searcher
```
for deep retrieval, or
```
ccf-experiment-designer
```
for experiment design. For publication-grade figure/table layout, visual contracts, palette choices, caption placement, or manuscript visual integration, hand off to
```
ccf-visual-composer
```
. Do not fill gaps by invention.
Before calling text ready, rerun
```
ccf-humanization
```
, verify that only confirmed full method versions appear, and run the final prose self-audit in
```
references/prose-quality-guardrails.md
```
. For a full section or paper, run
```
scripts/check_prose_quality.py
```
when local execution is available. Resolve all hard failures, especially a full-paper em-dash count above three. Keep any unresolved concern outside the manuscript as a user-review warning. Run a local score-risk check only when it adds concrete value; when scientific judgment is needed, hand off to
```
ccf-paper-reviewer
```
.

运行
```
ccf-humanization
```
前置步骤，然后确定模式、请求的输出格式、目标会议/期刊、论文类型、草拟状态、可用证据、目标篇幅/页数预算以及是否授权更改想法范围。对于源格式保留、模糊输出选择或从零开始的LaTeX草拟，加载
```
references/output-style-policy.md
```
。对于精确的JSON/文本模式，直接遵循用户指定的模式，无需加载该参考文件。
如果用户提供了现有稿件或标准草拟版本，除非用户要求新结构或快照，否则保留其格式并原地修订该文件。遵循
```
../ccf-common/references/artifact-contracts.md
```
：普通迭代无需创建
```
v2
```
、
```
revised
```
、带日期、带时间戳或
```
final-final
```
的副本。使用版本库历史记录进行回滚，避免将临时提示、评论和尝试性内容存入项目目录。
仅在需要时选择会议和范例。对于完整稿件、从零开始的投稿草拟或明确的风格/会议适配，加载会议指南以及
```
references/exemplars/index.md
```
中最多匹配的范例卡片。如果此类任务未指定会议，加载
```
references/custom-format/default-user-format.md
```
；仅将NeurIPS作为完整稿件LaTeX格式的备选。然后加载
```
references/length-budget-policy.md
```
并创建章节级篇幅规划。对于仅涉及有限证据的成果、精确格式响应、普通润色、标题、摘要或章节片段，除非用户明确要求会议/范例适配，否则跳过此步骤。
对于完整稿件、篇幅较大的引言、明确的故事线生成、叙事设计或研究洞察框架，使用
```
references/storyline-blueprint.md
```
构建或更新全局故事线。对于仅涉及有限证据的摘要或概述，使用此处已说明的紧凑结构，无需加载蓝图：科学背景 -> 知识缺口 -> 核心洞察 -> 方法机制 -> 证据 -> 严谨主张。此内容仅内部使用，除非用户要求规划。
准备引用：对于需要文献支持的完整稿件和章节草拟，加载
```
references/citation-workflow.md
```
，确定引用位置，获取已验证的条目，并更新项目参考文献。如果用户明确限制任务仅使用给定证据或禁止浏览/引用，需遵守该约束，不得加载引用工作流或搜索文献。如果需要相关竞品文献但未知，先交由
```
ccf-literature-searcher
```
处理，再继续后续工作。
始终应用
```
references/prose-quality-guardrails.md
```
。仅在处理完整论文、篇幅较大的章节草拟或明确的结构设计时，加载
```
references/research-writing-patterns.md
```
、
```
references/section-modules.md
```
和
```
references/writing-checklists.md
```
。默认情况下，仅涉及有限内容的摘要、方法概述、标题、段落编辑或精确JSON任务无需加载这些参考文件。对于完整稿件，将草拟版本与篇幅预算对比，按需扩充或压缩。
编译与检查（当支持LaTeX引擎时，完整稿件必须执行）：编译草拟版本，测量实际页数，并与步骤3中的会议预算对比。如果篇幅低于目标>15%，通过补充机制细节、科学相关方法说明、实验设置和实质性分析进行扩充，不得添加通用局限性、防御性文本、内部确认说明或编造结果。如果篇幅超过目标>10%，使用
```
references/compression-rules.md
```
进行压缩。每次重大修改后重新编译，重复此过程直至页数在容忍范围内。记录最终状态：
```
underfilled / target-fit / draft-over / final-over / not compiled
```
。
压缩时，加载
```
references/compression-rules.md
```
；以相同格式返回压缩后的文本，仅当用户要求或内容被大量移除时才包含删减日志。
演示文稿适配时，仅根据稿件和给定证据生成幻灯片/海报/演讲稿/问答内容，遵循用户要求的幻灯片/海报/脚本格式。
如果当前文献、基线或实验缺失，提出针对性问题或交由
```
ccf-literature-monitor
```
处理最新论文/竞品跟踪，交由
```
ccf-literature-searcher
```
处理深度检索，或交由
```
ccf-experiment-designer
```
处理实验设计。对于出版级别的图/表布局、视觉规范、配色选择、标题位置或稿件视觉整合，交由
```
ccf-visual-composer
```
处理。不得通过编造填补缺口。
在确认文本就绪前，重新运行
```
ccf-humanization
```
，验证仅出现已确认的完整方法版本，并执行
```
references/prose-quality-guardrails.md
```
中的最终文本自审。当支持本地执行时，对完整章节或论文运行
```
scripts/check_prose_quality.py
```
。解决所有严重问题，尤其是整篇论文中破折号数量超过三个的情况。将未解决的问题放在稿件外作为用户评审警告。仅当能带来具体价值时才进行本地风险评分检查；当需要科学判断时，交由
```
ccf-paper-reviewer
```
处理。

Post-Writing Coordination

写作后协调

After producing a complete manuscript, inform the user which CCFA skills are relevant to the next stage unless the user requested an exact output shape, JSON-only response, direct file edit, or no process commentary. Do not append handoff material that would violate the requested format.

State concisely what was produced and which skills can pick up from here:

``text Next CCFA skills available:

ccf-paper-reviewer: scientific review, score prediction, reviewer simulation, venue-fit check
ccf-integrity-auditor: citation existence, claim-evidence consistency, bibtex correctness
ccf-submission-checker: venue template compliance, page/anonymity limits, LaTeX compilation
ccf-experiment-designer: missing experiments, baseline design, ablation planning
ccf-visual-composer: publication-grade figures/tables, palette, caption, layout, visual QA
ccf-literature-monitor: recent-paper tracking, competitor monitoring, novelty-threat alerts
ccf-literature-searcher: missing related work, closest-competitor search
ccf-rebuttal-writer: reviewer response, rebuttal drafting, revision ledger
ccf-pipeline-orchestrator: full-project planning, stage gates, next-skill routing ``

Only list skills that are actually relevant to the paper current state. Do not suggest a skill that would have nothing to do.

生成完整稿件后，告知用户下一阶段相关的CCFA技能，除非用户要求精确输出格式、仅JSON响应、直接文件编辑或无需流程说明。不得添加违反请求格式的交接内容。

简洁说明已生成的内容以及后续可承接的技能：

``text Next CCFA skills available:

ccf-paper-reviewer: scientific review, score prediction, reviewer simulation, venue-fit check
ccf-integrity-auditor: citation existence, claim-evidence consistency, bibtex correctness
ccf-submission-checker: venue template compliance, page/anonymity limits, LaTeX compilation
ccf-experiment-designer: missing experiments, baseline design, ablation planning
ccf-visual-composer: publication-grade figures/tables, palette, caption, layout, visual QA
ccf-literature-monitor: recent-paper tracking, competitor monitoring, novelty-threat alerts
ccf-literature-searcher: missing related work, closest-competitor search
ccf-rebuttal-writer: reviewer response, rebuttal drafting, revision ledger
ccf-pipeline-orchestrator: full-project planning, stage gates, next-skill routing ``

仅列出与当前论文状态实际相关的技能。不得推荐无关技能。

Adaptive Output Contract

自适应输出协议

Polish/rewrite/compress existing text: output the revised text first, in the same format. Add brief notes only for changed meaning, unsupported claims, or user-requested rationale.
Draft from idea: output the requested artifact. For a manuscript request, default to a LaTeX draft with venue assumptions, page budget, and
```
TBD
```
evidence placeholders. If no target venue guide exists, use NeurIPS fallback. A short seed should become a full submission-shaped manuscript unless the user explicitly asked for a short demo or skeleton.
Section planning: output the section plan or paragraph roles only when the user asks for planning or the input is too incomplete to draft responsibly.
Presentation: output slides, poster copy, talk script, or Q&A in the requested format.
Full standard work may include a compact status block: mode, venue assumption, evidence gaps, and next CCFA owner.

润色/改写/压缩现有文本：首先输出相同格式的修订文本。仅当含义改变、主张无依据或用户要求说明理由时添加简短说明。
根据想法草拟：输出所需成果。对于稿件请求，默认生成带有会议假设、页数预算和
```
TBD
```
证据占位符的LaTeX草拟版本。如果无目标会议指南，使用NeurIPS备选方案。除非用户明确要求简短演示或框架，否则简短的初始想法应生成完整的投稿格式稿件。
章节规划：仅当用户要求规划或输入内容过于不完整而无法负责任地草拟时，输出章节规划或段落作用说明。
演示文稿：按照用户要求的格式输出幻灯片、海报文案、演讲稿或问答内容。
完整标准工作可包含紧凑状态块：模式、会议假设、证据缺口以及下一任CCFA负责人。

Reference Files

参考文件

references/venue-guides/index.md

and

references/venue-guides/<venue>.md

: venue writing and format constraints.

```
references/output-style-policy.md
```
: user-format priority, edit-format preservation, and NeurIPS fallback behavior.
```
references/length-budget-policy.md
```
: venue page/word budget, underfilled draft expansion, compile-adjust loop, and compression trigger.
```
references/research-writing-patterns.md
```
: section-level patterns, dense output rules, and exemplar-mode adaptation.
```
references/prose-quality-guardrails.md
```
: anti-defensive prose rules, paragraph cohesion, terminology consistency, evidence-bounded claims, sentence rhythm, and final prose self-audit.
```
scripts/check_prose_quality.py
```
: deterministic, non-mutating prose-pattern check for em dashes, opening filler, repetitive structure, punctuation density, and sentence rhythm.
```
references/storyline-blueprint.md
```
: whole-paper story, multi-expert storyline generation, scientific storytelling structure, and claim construction.
```
references/section-modules.md
```
: section drafting.
```
references/writing-checklists.md
```
: readiness checks.
```
references/score-lifting-loop.md
```
: apply evidence-backed reviewer deductions without creating a second score.
```
references/expert-review-loop.md
```
: compatibility bridge from reviewer findings to writing-owned edits.
```
../ccf-common/references/review-output-standards.md
```
: quantitative feedback, reviewer-panel discipline, score-change conditions, and final output self-check.
```
references/compression-rules.md
```
: page/word compression.
```
references/table-style-guide.md
```
: LaTeX table beautification: booktabs rules, number precision, narrow-column fixes, caption style, and placement.
```
references/exemplars/
```
: style-move references; never copy content.
```
../ccf-humanization/references/humanization-policy.md
```
: priority rules for direct academic prose, warning-only issues, non-injection, unfavorable information, and generic SHA-256/checksum removal.
```
../ccf-humanization/references/experiment-discipline.md
```
: confirmed full method identities and the prohibition on simplified publication variants.

references/venue-guides/index.md

和

references/venue-guides/<venue>.md

：会议写作与格式约束。

```
references/output-style-policy.md
```
：用户格式优先级、编辑格式保留以及NeurIPS备选规则。
```
references/length-budget-policy.md
```
：会议页数/字数预算、草拟篇幅不足时的扩充、编译调整循环以及压缩触发条件。
```
references/research-writing-patterns.md
```
：章节级模式、高密度输出规则以及范例模式适配。
```
references/prose-quality-guardrails.md
```
：反防御性文本规则、段落连贯性、术语一致性、主张有证据支撑、句式节奏以及最终文本自审。
```
scripts/check_prose_quality.py
```
：针对破折号、开头填充语、重复结构、标点密度和句式节奏的确定性、非修改性文本模式检查工具。
```
references/storyline-blueprint.md
```
：整篇论文的故事线、多专家故事线生成、科学叙事结构以及主张构建。
```
references/section-modules.md
```
：章节草拟。
```
references/writing-checklists.md
```
：就绪检查清单。
```
references/score-lifting-loop.md
```
：应用基于证据的评审扣分规则，无需生成第二个分数。
```
references/expert-review-loop.md
```
：评审结果与写作修改之间的兼容衔接机制。
```
../ccf-common/references/review-output-standards.md
```
：量化反馈、评审小组规范、分数变更条件以及最终输出自审。
```
references/compression-rules.md
```
：页数/字数压缩规则。
```
references/table-style-guide.md
```
：LaTeX表格美化规则：booktabs规范、数字精度、窄列修复、标题样式以及放置位置。
```
references/exemplars/
```
：风格参考；不得复制内容。
```
../ccf-humanization/references/humanization-policy.md
```
：直接学术文本的优先级规则、仅警告类问题、不插入无关内容、不利信息处理以及通用SHA-256/校验和移除规则。
```
../ccf-humanization/references/experiment-discipline.md
```
：已确认的完整方法标识以及禁止使用简化出版变体的规则。