Install any skill in seconds. Free to start, no credit card required.
Get Started Free →本技能专门处理 **Category C:排版一致性缺陷**,包括:
.claude/skills/openraiser-consistency-polisher-skill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 27% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 203% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 110% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 191% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 225% | 0% |
本技能专门处理 Category C:排版一致性缺陷,包括:
\resizebox 滥用导致字号参差)该技能由 code-surgeon-agent 调用,通过对表格字号、图片尺寸策略、标题格式的统一化处理,消除全篇排版风格割裂,使论文呈现出专业、一致的视觉形象。
一致性缺陷通常不致命,但严重影响审稿人对论文专业度的第一印象。本技能的目标是建立并强制执行全篇统一的排版规范。
| 缺陷 ID | 描述 | 优先级 | 是否自动修复 | |---------|------|--------|-------------| | C1 | 表格字号不统一 | Medium | 是 | | C2 | 图片分辨率/风格不一致 | Low | 否(仅诊断并建议) | | C3 | Caption 格式不统一 | Medium | 是 |
| 输入项 | 来源 | 说明 | |--------|------|------| | 主 .tex 文件路径 | 项目上下文 | 需修改的源文件 | | 排版侦探报告 | layout-detective-agent 输出 | 包含 C 类缺陷的具体描述 | | 模板信息 | templates.yaml 或上下文 | 会议/期刊模板的字号、栏宽等默认参数 |
json{ "skill": "consistency-polisher", "status": "success | partial | failed", "modified_files": ["main.tex", "tables/experiments.tex"], "changes": [ { "defect_id": "C1", "object": "Table 2", "action": "移除 \\resizebox,改用 tabularx 并统一使用 \\small", "before_snippet": "\\resizebox{\\linewidth}{!}{\\begin{tabular}{...}", "after_snippet": "{\\small\\begin{tabularx}{\\linewidth}{...}}" } ], "suggestions": [ { "defect_id": "C2", "object": "Figure 5", "message": "图片分辨率过低(72 DPI),建议替换为至少 300 DPI 的矢量图或高分辨率位图" } ], "unresolved": [] }
问题特征:
\small,另一个 \tiny)。\resizebox{\linewidth}{!}{...} 整体缩放表格,导致文字被非线性缩放而变形,且与其他表格字号不一致。修复策略(按优先级):
\resizebox,改用 tabularx 自适应宽度\resizebox 是表格一致性的头号杀手,必须无条件移除。
latex% 修改前 \begin{table} \centering \resizebox{\linewidth}{!}{ \begin{tabular}{|l|c|c|c|} \hline Method & Metric1 & Metric2 & Metric3 \\ \hline Ours & 95.2 & 87.3 & 78.1 \\ \hline \end{tabular} } \caption{Results} \end{table} % 修改后 \begin{table} \centering \small % 统一字号 \begin{tabularx}{\linewidth}{|l|X|X|X|} \hline Method & Metric1 & Metric2 & Metric3 \\ \hline Ours & 95.2 & 87.3 & 78.1 \\ \hline \end{tabularx} \caption{Results} \end{table}
注意:确保导言区已加载 tabularx 宏包。若未加载,需添加 \usepackage{tabularx}。
检查全篇表格的字号设置,确定一个“锚点字号”(通常是模板默认的正文字号,或 \small),将所有表格统一为该字号。
识别当前字号:
\begin{table} 后的 \small、\footnotesize、\scriptsize、\tiny 等命令。\resizebox 并评估其实际视觉字号。统一方法:
\small;若仍超宽,考虑改为 tabularx 而非进一步缩小字号。\tiny 或 \scriptsize(不可阅读)。latex% 统一风格示例:在每个 table 环境内首行添加 \small \begin{table} \small \centering \begin{tabularx}{\linewidth}{...} ... \end{tabularx} \end{table}
有时作者使用超小字号是因为表格列太多、内容太宽。此时应优先重构列格式(如合并相似列、使用缩写表头、改为跨栏表),而非暴力缩小字号。
latex% 修改前(滥用 \tiny) \begin{table} \tiny \begin{tabular}{|l|c|c|c|c|c|c|c|c|} ... \end{tabular} \end{table} % 修改后(合并为关键列 + 缩写表头 + 跨栏宽表) \begin{table*} \small \begin{tabularx}{\textwidth}{|l|X|X|X|} \hline \textbf{Method} & \textbf{Prec.} & \textbf{Rec.} & \textbf{F1} \\ \hline ... \end{tabularx} \end{table*}
问题特征:
.pdf、.png、.jpg,且质量参差。本技能的定位:仅诊断和建议,不自动修改图片内容。
.pdf(矢量图)、.eps(矢量图)。.png(≥300 DPI)。.jpg、屏幕截图。json{ "defect_id": "C2", "object": "Figure 5 (figures/ablation.png)", "issue": "图片分辨率约 72 DPI,放大后锯齿明显", "suggestion": "请使用原始矢量图导出为 PDF,或以至少 300 DPI 重新渲染位图" }
json"suggestions": [ { "defect_id": "C2", "object": "Figure 1, Figure 3, Figure 5", "message": "三张图使用了三种不同的配色风格,建议统一使用会议模板推荐的配色方案。" } ]
问题特征:
修复策略:
caption 宏包配置在导言区加载 caption 宏包并设置全局样式,一次性统一全篇标题格式。
latex\usepackage{caption} \captionsetup{ font=small, % 字号 labelfont=bf, % 标签加粗("Figure 1:" 中的 "Figure 1" 部分) textfont=it, % 标题文本斜体(可选) labelsep=period, % 标签与文本分隔符:period = 句点,colon = 冒号 justification=raggedright, % 左对齐(或 justified 两端对齐) singlelinecheck=false % 即使单行也应用设置 }
针对表格和图片的差异化设置:
latex% 表格标题默认在上方,图片在下方,无需额外设置,LaTeX 自动处理。 % 若需微调间距: \captionsetup[table]{position=above, skip=6pt} \captionsetup[figure]{position=below, skip=6pt}
搜索 .tex 文件中的 \caption 命令,移除其前后的手动字体设置(如 \textbf{\caption{...}}),因为这些手动设置会覆盖全局配置。
latex% 修改前 \caption{\textbf{This is a bold caption.}} % 修改后(依赖全局 caption 设置) \caption{This is a caption with unified style.}
检查所有 \caption{} 的内容,确保标点风格一致。推荐:
latex% 统一移除标题末尾的句号 \caption{Results on the validation set.} → \caption{Results on the validation set}
LaTeX 中表格标题应放在 \begin{tabular} 之前,图片标题应放在 \includegraphics 之后。
latex% 正确的表格标题位置 \begin{table} \centering \caption{Table caption here} % 在上方 \begin{tabular}{...} ... \end{tabular} \end{table} % 错误的表格标题位置(标题在表格下方) \begin{table} \centering \begin{tabular}{...} ... \end{tabular} \caption{This caption is misplaced.} \end{table}
若发现位置错误,将 \caption 命令移动到正确位置。
在修复 C 类缺陷时,应同时对全篇进行以下检查:
| 检查项 | 目标状态 | 修复方法 | |--------|----------|----------| | 所有表格是否使用相同字号? | 全篇统一(通常 \small 或模板默认) | 移除 \resizebox,添加统一的字号命令 | | 所有图片是否使用 \linewidth? | 是(单栏)或 \textwidth(跨栏) | 替换固定宽度为相对宽度 | | 所有 Caption 字体是否一致? | 是 | 使用 caption 宏包全局配置 | | 表格标题是否都在表格上方? | 是 | 移动 \caption 命令 | | 图片标题是否都在图片下方? | 是 | 移动 \caption 命令 | | 标题结尾标点是否统一? | 全篇统一(推荐无句号) | 移除或添加标点 | | 是否使用了 booktabs 绘制三线表? | 推荐 | 将 \hline 替换为 \toprule、\midrule、\bottomrule |
\resizebox 移除常与 D1 的表格溢出修复联动,可合并处理。每完成一项修复后:
\resizebox 而超宽),若有,转交 overflow-repair 处理。Consistency Polisher Skill 就绪。 等待调用,消除排版风格割裂,交付视觉统一的作品。
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | fail→fail | 10,032 | 6,901 | -31% | 1 | 1 | 0% | 1,547 | 4,232 | +174% | 0 | 0 | — |
case-01 | fail→pass | 19,621 | 10,347 | -47% | 1 | 1 | 0% | 4,065 | 5,178 | +27% | 0 | 0 | — |
case-02 | fail→pass | 11,593 | 11,288 | -3% | 1 | 1 | 0% | 1,821 | 5,525 | +203% | 0 | 0 | — |
case-03 | fail→fail | 13,645 | 8,250 | -40% | 1 | 1 | 0% | 2,696 | 4,725 | +75% | 0 | 0 | — |
case-04 | pass→pass | 13,371 | 5,884 | -56% | 1 | 1 | 0% | 2,318 | 4,162 | +80% | 0 | 0 | — |
case-05 | pass→pass | 11,554 | 10,966 | -5% | 1 | 1 | 0% | 1,849 | 4,945 | +167% | 0 | 0 | — |
case-06 | fail→pass | 11,444 | 4,922 | -57% | 1 | 1 | 0% | 1,853 | 3,887 | +110% | 0 | 0 | — |
case-07 | pass→pass | 10,172 | 7,582 | -25% | 1 | 1 | 0% | 1,566 | 4,359 | +178% | 0 | 0 | — |
case-08 | fail→pass | 8,371 | 4,761 | -43% | 1 | 1 | 0% | 1,337 | 3,894 | +191% | 0 | 0 | — |
case-09 | pass→pass | 10,726 | 7,390 | -31% | 1 | 1 | 0% | 1,802 | 4,429 | +146% | 0 | 0 | — |
case-10 | fail→fail | 4,095 | 3,745 | -9% | 1 | 1 | 0% | 596 | 3,667 | +515% | 0 | 0 | — |
case-12 | fail→pass | 7,396 | 5,122 | -31% | 1 | 1 | 0% | 1,217 | 3,958 | +225% | 0 | 0 | — |
case-13 | pass→pass | 7,285 | 5,710 | -22% | 1 | 1 | 0% | 1,197 | 4,164 | +248% | 0 | 0 | — |
case-14 | pass→pass | 8,422 | 8,339 | -1% | 1 | 1 | 0% | 1,436 | 4,420 | +208% | 0 | 0 | — |
case-15 | fail→pass | 10,310 | 3,780 | -63% | 1 | 1 | 0% | 1,693 | 3,877 | +129% | 0 | 0 | — |
case-16 | fail→pass | 10,541 | 4,212 | -60% | 1 | 1 | 0% | 1,592 | 3,815 | +140% | 0 | 0 | — |
case-17 | pass→pass | 3,393 | 3,644 | +7% | 1 | 1 | 0% | 537 | 3,686 | +586% | 0 | 0 | — |
case-18 | pass→pass | 5,004 | 5,510 | +10% | 1 | 1 | 0% | 908 | 4,048 | +346% | 0 | 0 | — |
case-19 | pass→pass | 5,109 | 7,261 | +42% | 1 | 1 | 0% | 829 | 4,305 | +419% | 0 | 0 | — |
case-20 | pass→pass | 15,228 | 12,098 | -21% | 1 | 1 | 0% | 2,502 | 5,127 | +105% | 0 | 0 | — |
case-21 | pass→fail | 14,527 | 11,535 | -21% | 1 | 1 | 0% | 2,359 | 4,915 | +108% | 0 | 0 | — |
case-22 | fail→fail | 12,528 | 11,510 | -8% | 1 | 1 | 0% | 2,316 | 5,068 | +119% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +27 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.