A code-driven presentation generation framework. 像构建软件工程一样生成演示文稿。
$ npx -y skills add sunbigfly/ppt-agent-skills --agent claude-code
Run the curl in your terminal, the rest in Claude Code.
What's inside
PPT Agent 以严格的状态机驱动多 Agent 协作,将一句话需求输出为专业级 PPTX 文件,从根源解决传统大模型生成的幻觉、重叠与布局混乱问题。
2026-04-09 · v4.1
density_bias 与 Step 4 density_label / density_contract 合同,页面密度预算开始成为明确的可校验约束。visual_qa.py 新增 planning + html 双层断言,除了看图,也会核对结构与装饰预算。subagent_logger.py,把 PageAgent / PagePatchAgent 的阶段命令与运行日志补齐,断点返工更容易审计与回放。npx skills add sunbigfly/ppt-agent-skills
Subagent 阶段隔离:Research / Outline / Style / Planning 四大阶段各自运行独立的子代理,Context 不互染。每个子代理创建时强制携带 SUBAGENT_MODEL 参数,禁止走默认回退。
像素级 Visual QA 闭环:每页 HTML 构建后自动截图,由大模型进行视觉审计。检测到布局溢出后,子代理以 DOM + CSS 结构重写的方式消除冲突,而非依赖间距微调。
无状态断点恢复:全流程不依赖任何进度状态文件。中断后通过扫描磁盘上已存在的产物文件(outline.txt / style.json / slide-N.png 等)自动推断恢复点。
数据层与渲染层隔离:每页先生成并由 planning_validator.py 通过校验的 JSON 合同,再驱动 HTML 渲染。针对生成的静态文件拥有 html 级物理结构探测器拦截非标骨架。
《阴阳割线》设计哲学:首创在底层封死绝对的物理安全墙与基准坐标流(阴)的前提下,向 AI 模型全量释放针对 Typography、重叠深度与留白排版的特权(阳),告别传统的平庸公文感。
极速底层光栅化引擎:内置的 Puppeteer 引擎彻底摒弃硬延时,全面接入事件全息钩子 (document.fonts.ready 及全节点监听),实现 0 丢字体、0 秒白等的极速快门截图与 SVG 源生解析。
双引擎 PPTX 导出:PNG 光栅流保证跨平台 100% 视觉还原;SVG 矢量流保留字体可独立编辑。
P0 采访 → P1 分支确认
P2A 联网检索 / P2B 本地资料压缩
P3 叙事大纲 → P3.5 全局风格锁定
P4 逐页并行生产(Planning → HTML → Visual QA)
P5 Preview + 双 PPTX 导出
每个阶段产物落盘后经 Gate 校验放行,失败只回退当前步,不影响其他页进度。
interview-qa.txt → requirements-interview.txt
→ search.txt + search-brief.txt | source-brief.txt
→ outline.txt → style.json
→ planningN.json → slide-N.html → slide-N.png
→ preview.html → presentation-{png,svg}.pptx → delivery-manifest.json
本项目以 Agent Skill 形式运行,无需独立部署。在支持 Skill 的代理环境中直接输入需求即可触发完整流程:
"帮我生成一份关于 2026 年具身智能发展趋势的 15 页路演 Deck,暗色科技风格。"
所有产物输出至 ppt-output/runs/<RUN_ID>/,包含网页预览和双格式 PPTX。
ppt-agent-skill/
├── SKILL.md # 主控制台:状态机、Gate、恢复规则
├── scripts/ # 执行脚本(validator / harness / exporter)
├── references/ # 按需挂载的 Markdown 知识源
│ ├── playbooks/ # 各阶段子代理执行手册
│ ├── styles/ # 主题风格规范
│ ├── layouts/ # 版式资源
│ ├── charts/ # 图表模板
│ └── blocks/ # UI 组件
└── assets/
已链接认可 LINUX DO 社区 的友情链接。
.gitignore
assets/
logo.png
screenshots/
slide1.png
slide2.png
slide3.png
slide4.png
LICENSE
README_EN.md
README.md
references/
blocks/
card-styles.md
comparison.md
diagram.md
image-hero.md
matrix-chart.md
people.md
quote.md
README.md
timeline.md
charts/
comparison-bar.md
funnel.md
kpi.md
metric-row.md
progress-bar.md
radar.md
rating.md
README.md
ring.md
runtime-chart-rules.md
sparkline.md
stacked-bar.md
timeline.md
treemap.md
waffle.md
cli-cheatsheet.md
design-runtime/
css-weapons.md
data-type-decoration-mapping.md
data-type-visual-mapping.md
design-specs.md
director-command-examples.md
director-command-rules.md
layouts/
asymmetric.md
hero-top.md
l-shape.md
mixed-grid.md
primary-secondary.md
README.md
single-focus.md
symmetric.md
t-shape.md
three-column.md
waterfall.md
page-templates/
cover.md
end.md
section.md
toc.md
playbooks/
outline-phase1-playbook.md
outline-phase2-playbook.md
research-phase1-playbook.md
research-phase2-playbook.md
source-phase1-playbook.md
source-phase2-playbook.md
step4/
page-html-playbook.md
page-planning-playbook.md
page-review-playbook.md
style-phase1-playbook.md
style-phase2-playbook.md
principles/
cognitive-load.md
color-psychology.md
composition.md
data-visualization.md
design-principles-cheatsheet.md
narrative-arc.md
README.md
runtime-failure-modes.md
visual-hierarchy.md
prompts/
module-structured-interview-ui.md
module-text-interview-fallback.md
step4/
tpl-page-audit-request.md
tpl-page-breakpoint-orchestrator.md
tpl-page-html.md
tpl-page-orchestrator.md
tpl-page-planning.md
tpl-page-review.md
tpl-interview-structured-ui.md
tpl-interview-text-fallback.md
tpl-interview.md
tpl-outline-orchestrator.md
tpl-outline-phase1.md
tpl-outline-phase2.md
tpl-research-synth-orchestrator.md
tpl-research-synth-phase1.md
tpl-research-synth-phase2.md
tpl-source-synth-orchestrator.md
tpl-source-synth-phase1.md
tpl-source-synth-phase2.md
tpl-style-orchestrator.md
tpl-style-phase1.md
tpl-style-phase2.md
README.md
styles/
blue-white.md
dark-tech.md
fresh-green.md
luxury-purple.md
minimal-gray.md
README.md
royal-red.md
runtime-style-palette-index.md
runtime-style-rules.md
vibrant-rainbow.md
warm-earth.md
scripts/
check_skill.py
contract_validator.py
html_packager.py
html2png.py
html2svg.py
milestone_check.py
planning_validator.py
png2pptx.py
prompt_harness.py
README.md
resource_loader.py
smoke_skill.py
subagent_logger.py
svg2pptx.py
visual_qa.py
workflow_versions.py
SKILL.mdFAQ
ppt-agent-skills is a Claude Code plugin with 1 hand-picked skill for content work, indexed on Flowy. Install it with the command on its page. It includes ppt-agent-skills. Its skills do not fire on their own yet. Request auto-invocation to have Flowy route them as you prompt. Free and open source.