Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when rebuilding slide screenshots, image-only PPTX files, AI-generated reference slides, or user-edited PowerPoint drafts into editable PPTX deliverables.
.claude/skills/kiakun-collab-ppt-rebuild-workflow/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-13 | ✗→✓ | ▲ Improved | 66% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 142% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -1% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 54% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 33% | 0% |
把参考画面重构成符合用户编辑需求的 PPTX。默认忠实重构，先构建可用初版，再根据实际问题定向校准。
只发一张图并说“转 PPT / 重构”：默认 Mode B + balanced + Level 2，文字、表格与主要结构可编辑，复杂主视觉可合成图片。直接说明边界并开始，不要求用户填写模式、QA、字体、输出名。只有真实需求冲突或关键内容缺失才提问。
rebuild_workflow.py init 固定路由，再启动构建；不在交付前补跑。layout-spec 驱动构建，见 data-driven-build.md。公共样式和模板定义一次，每页只写内容与差异；展开布局与 QA 兼容文件由程序生成，不重复手填。按 visual-extraction-pass.md 保留实际观测，不倒填测量。visual-extraction.shapes[] 并保留来源。| 需求 | 文件 | | --- | --- | | A 整页图预览 | mode-selection.md；Level 1 | | B 文字及主要结构可编辑 | semi-editable-workflow.md | | C 主视觉也独立分层 | full-layered-workflow.md | | D 用户要求新视觉方向 | mode-d-workflow.md | | E 修改用户当前 PPTX | incremental-edit-workflow.md | | 新建重构页、统一数据或批量模板复用 | data-driven-build.md：统一坐标、模板合并和 Presentations 适配器 | | 没有干净底图、需要抠图/分层 | asset-policy.md | | 复杂背景中的人物/产品需保留原貌并独立 | 辅助分割方法：粗轮廓 + GrabCut，限定局部修正次数 | | 高精度要求或已发现位置/字体偏差 | autonomous-calibration.md | | 复杂光晕、渐隐、图片融合 | visual-transition-strategy.md | | 构建细节/异常 | implementation-guardrails.md | | 具体视觉疑难 | visual-overlap-qa.md、visual-fidelity-qa.md |
不要开场加载全部模式、QA 子文档和模板。同一任务的派生记录从同一份抽取数据生成，避免重复填表。
实际制作读 runtime-integration.md。Presentations 激活时构建和渲染遵守其当前契约，本 Skill 补充边界与 QA，不重复视觉评分。其他重构 skill 的占位图、强制询问字体等默认值不得覆盖已确定的用户需求。
普通权限下先创建输出目录，使用新版本文件；先保存 PPTX 再渲染。预览失败时继续包内 QA，报告视觉验收未完成。保留用户当前文件；权限与清理遵循当前线程规则。
命令与字段见 script-output-contracts.md。
rebuild_workflow.py 是日常入口：init / preflight / assets / bind-copy / run / prepare-review / report。任务记录用 task-input-template.json；审核填写一份 prepare-review 生成的待审文件，再用 evidence-input-template.json 汇总。完整 qa-report 模板仅作字段参考，不逐项重复手写。A/E 保留专用验证；Level 2/3 检查表位于 assets/templates/。
仅在平台可用且已获授权时使用子 agent。单页默认不拆多个重复看图的审计者，独立工作参考 subagent-prompts.md。
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 25,503 | 9,955 | -61% | 1 | 1 | 0% | 4,669 | 3,342 | -28% | 0 | 0 | — |
case-12 | pass→pass | 18,363 | 15,652 | -15% | 1 | 1 | 0% | 1,811 | 3,140 | +73% | 0 | 0 | — |
case-13 | fail→pass | 19,895 | 16,052 | -19% | 1 | 1 | 0% | 2,052 | 3,397 | +66% | 0 | 0 | — |
case-01 | fail→fail | 14,528 | 10,683 | -26% | 1 | 1 | 0% | 1,404 | 2,432 | +73% | 0 | 0 | — |
case-02 | fail→pass | 17,104 | 60,959 | +256% | 1 | 1 | 0% | 1,593 | 3,863 | +142% | 0 | 0 | — |
case-03 | fail→fail | 21,278 | 27,822 | +31% | 1 | 1 | 0% | 3,618 | 5,305 | +47% | 0 | 0 | — |
case-04 | fail→fail | 19,343 | 28,743 | +49% | 1 | 1 | 0% | 2,104 | 5,510 | +162% | 0 | 0 | — |
case-05 | fail→fail | 35,829 | 51,290 | +43% | 1 | 1 | 0% | 4,605 | 7,473 | +62% | 0 | 0 | — |
case-07 | pass→pass | 20,713 | 10,430 | -50% | 1 | 1 | 0% | 1,923 | 3,308 | +72% | 0 | 0 | — |
case-08 | pass→pass | 22,136 | 14,778 | -33% | 1 | 1 | 0% | 2,360 | 3,348 | +42% | 0 | 0 | — |
case-09 | fail→pass | 23,209 | 8,872 | -62% | 1 | 1 | 0% | 3,341 | 3,311 | -1% | 0 | 0 | — |
case-10 | fail→pass | 10,757 | 8,206 | -24% | 1 | 1 | 0% | 1,528 | 2,352 | +54% | 0 | 0 | — |
case-11 | fail→pass | 23,122 | 7,637 | -67% | 1 | 1 | 0% | 2,287 | 3,049 | +33% | 0 | 0 | — |
case-14 | fail→pass | 18,230 | 13,429 | -26% | 1 | 1 | 0% | 1,843 | 2,530 | +37% | 0 | 0 | — |
case-15 | fail→pass | 11,943 | 4,048 | -66% | 1 | 1 | 0% | 1,682 | 2,370 | +41% | 0 | 0 | — |
case-16 | fail→pass | 14,507 | 12,056 | -17% | 1 | 1 | 0% | 2,213 | 2,916 | +32% | 0 | 0 | — |
case-17 | fail→pass | 20,839 | 9,862 | -53% | 1 | 1 | 0% | 2,325 | 3,237 | +39% | 0 | 0 | — |
case-18 | pass→pass | 18,950 | 11,215 | -41% | 1 | 1 | 0% | 1,950 | 2,741 | +41% | 0 | 0 | — |
case-19 | fail→pass | 34,292 | 21,859 | -36% | 1 | 1 | 0% | 2,274 | 2,852 | +25% | 0 | 0 | — |
case-20 | fail→pass | 13,246 | 10,773 | -19% | 1 | 1 | 0% | 1,885 | 2,777 | +47% | 0 | 0 | — |
case-21 | fail→pass | 18,664 | 18,715 | +0% | 1 | 1 | 0% | 1,823 | 2,700 | +48% | 0 | 0 | — |
case-22 | fail→fail | 19,819 | 11,141 | -44% | 1 | 1 | 0% | 2,279 | 2,457 | +8% | 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 +55 percentage points is the difference between those two pass rates over the 22 comparable cases.
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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/6/2026 | +41% |
Other measured skills in the registry, with their headline benchmark lift.