1、不依赖任何第三方 PPT 服务； 2、给可以自行上传ppt模板和内容资料，约束让大模型根据上传的内容生成最终的成品 3.自动导出生成.pptx格式的文件

Generate professional .pptx presentations from any topic, uploaded documents, or reference materials. Supports custom PPT templates (filled via python-pptx, not XML editing). Works with any LLM configured in OpenClaw. Real chart rendering, smart content chunking, guided conversation flow.

YuchenTsai-design@yuchentsai-design

Install

openclaw skills install @yuchentsai-design/ppt-generator-skill

PPT Generator Skill

Generate polished .pptx presentations using the LLM configured in OpenClaw.
No external PPT API required — everything runs locally.

Phase 0 — Environment Check (First-Time Setup)

Run once after installing the skill, before anything else.

bash

python3 scripts/check_env.py

The script checks all required and optional dependencies and prints a clear report:

text

══════════════════════════════════════════════════════
  PPT Generator Skill — Environment Check
══════════════════════════════════════════════════════
  ✅  Python 3.9+      (3.11.4)
  ✅  Node.js          (v20.11.0)
  ✅  npm              (10.2.4)
  ✅  pptxgenjs        (installed)
  ✅  markitdown       (0.4.1)
  ✅  pypdf            (4.2.0)
  ✅  python-pptx      (0.6.23)
  ⚠️   LibreOffice      (not found)
  ⚠️   pdftoppm         (not found)
──────────────────────────────────────────────────────

  ⚠️  OPTIONAL — QA features will be unavailable:

  ▸ LibreOffice  — Optional, needed for QA visual inspection
      $ sudo apt install libreoffice       # Ubuntu/Debian
      $ brew install --cask libreoffice    # macOS

  ▸ pdftoppm  — Optional, needed for QA visual inspection
      $ sudo apt install poppler-utils     # Ubuntu/Debian
      $ brew install poppler               # macOS
══════════════════════════════════════════════════════

If required (❌) items are missing, exact platform-specific install commands are shown. Fix all required items before proceeding.

Auto-install missing Python packages:

bash

python3 scripts/check_env.py --fix

Machine-readable output (for CI / automated setup):

bash

python3 scripts/check_env.py --json

Phase 1 — Requirements Gathering

Do not run any scripts yet. First collect everything needed through a short conversation.

Fast-path — Skip questions when info is already provided

Before asking questions, check whether the user's first message already contains sufficient information. If at least two of the following can be inferred from their message, skip the corresponding question steps and proceed directly to Phase 2:

Info needed	Example signals	Maps to
Purpose/topic	"AI 趋势", "年度总结", "产品发布"	`{purpose}`, `{raw_topic}`
Content source	Uploaded files, or detailed description	`{uploaded_files}` or `{raw_topic}`
Slide count or style	"10 页", "科技风", "简约风"	`{slide_count}`, `{style}`

Example: User says "帮我做一个 10 页的 AI 趋势科技风 PPT" → purpose=科技主题, raw_topic=AI 趋势, slide_count=10, style=未来科技. Skip to Phase 2 style confirmation with pre-filled values.

If information is insufficient, proceed with the question steps below.

Step 1.1 — Understand purpose and topic

Ask the user:

请问这个 PPT 的主题和用途是什么？比如：

📚 学术报告 / 论文答辩

📊 工作汇报 / 季度总结 / 年终总结

🎓 培训课程 / 教学课件

💼 商业提案 / 投资路演

🚀 产品发布 / 项目介绍

🎨 活动策划 / 创意方案

其他（请描述）

Record the answer as {purpose}.

Step 1.2 — Collect content and reference materials

Ask the user:

请提供这次 PPT 的具体内容或参考资料，可以是：

直接描述主题要点（文字说明）

上传文档（支持 .txt / .md / .pdf / .docx / .pptx / .csv / .json）

两者都有也没问题，内容越详细，PPT 质量越高

Wait for the user's response:

Text description → record as {raw_topic}
Uploaded files → record as {uploaded_files}
Both → use both

Step 1.3 — Ask about PPT template

Ask the user:

您有现成的 PPT 模板文件吗？

有 → 请上传（.pptx 格式）

没有 → 我来为您推荐合适的风格

If a .pptx file is uploaded, ask a follow-up:

请问这个 .pptx 文件的用途是：

🎨 模板文件 — 保留它的视觉风格（背景、配色、字体、版式），填入新内容

📄 参考内容 — 提取其中的文字内容作为素材，视觉风格另行选择

Record the answer as {pptx_role}: "template" or "content".

Phase 2 — Style and Structure

If no template: suggest style (Step 2.1)

Run style analysis from what was collected in Phase 1:

bash

# Topic description only:
python3 scripts/suggest_style.py --query "{raw_topic}"

# Uploaded files only:
python3 scripts/analyze_content.py {uploaded_files} > content_analysis.json
python3 scripts/suggest_style.py --content-file content_analysis.json

# Both:
python3 scripts/analyze_content.py {uploaded_files} > content_analysis.json
python3 scripts/suggest_style.py --query "{raw_topic}" --content-file content_analysis.json

Present the suggestion to the user:

根据您的需求，推荐方案如下：

项目建议
风格分类 {suggested_category}
预计页数 {slide_count_estimate} 页
建议章节 {key_sections（逗号分隔）}

配色方案（3 选 1）：

# 方案名主色辅色强调色
1 {palette_candidates[0].name} ⬛ {primary} ⬜ {secondary} ◻ {accent}
2 {palette_candidates[1].name} ⬛ {primary} ⬜ {secondary} ◻ {accent}
3 {palette_candidates[2].name} ⬛ {primary} ⬜ {secondary} ◻ {accent}

字体搭配： 标题 {font_heading} / 正文 {font_body}

请选择配色方案（输入 1/2/3），或直接确认使用推荐方案 1。也可以告诉我需要调整的地方（换风格、增减页数等）。

项目	建议
风格分类	{suggested_category}
预计页数	{slide_count_estimate} 页
建议章节	{key_sections（逗号分隔）}

#	方案名	主色	辅色	强调色
1	{palette_candidates[0].name}	⬛ {primary}	⬜ {secondary}	◻ {accent}
2	{palette_candidates[1].name}	⬛ {primary}	⬜ {secondary}	◻ {accent}
3	{palette_candidates[2].name}	⬛ {primary}	⬜ {secondary}	◻ {accent}

Wait for confirmation. Record the chosen palette as {palette_hex} and update {slide_count} / {style} per feedback.

Important: After confirmation, add "style_category": "{suggested_category}" to the slide plan JSON (Step 3.2). This field controls which visual layout renderer is used during PPTX generation — each category produces a visually distinct slide structure.

Style categories quick reference:

分类	适合场景
企业商务	商业提案、投资路演、季度汇报
未来科技	AI/科技主题、数字化转型
年终总结	年度/季度/工作总结
扁平简约	数据报告、产品设计
中国风	传统文化、国学、非遗
卡通手绘	教学课件、培训、儿童内容
文艺清新	生活方式、旅行
文化艺术	展览、人文、艺术项目
创意趣味	头脑风暴、工作坊
学术研究	论文答辩、学术报告

If template uploaded (pptx_role = "template"): analyze it (Step 2.2)

bash

python3 scripts/analyze_template.py uploaded_template.pptx

Inform the user:

已分析您的模板文件，检测到：

配色：primary={primary}, secondary={secondary}

字体：标题={heading_font}，正文={body_font}

可用版式（{n} 种）：{layout_names}

我将保留这套视觉风格，只需您确认大致的页数即可。

Phase 3 — Content Analysis and PPT Generation

All information is now confirmed. Run scripts in sequence without further interruption.

Step 3.1 — Analyze content files

bash

python3 scripts/analyze_content.py {content_files} > content_analysis.json

# For large docs (>40k chars):
python3 scripts/analyze_content.py large_doc.pdf --max-chars 80000 > content_analysis.json

# Structure-only mode for very long docs:
python3 scripts/analyze_content.py large_doc.pdf --summary-only > content_analysis.json

If truncated_files is non-empty in the output, inform the user and offer --max-chars adjustment.

Step 3.1b — Content sizing for LLM (token optimization)

Before sending content to the LLM, apply these rules to control token usage:

Content size	Strategy	Expected tokens
< 5,000 chars	Send full content as-is	~1.5k
5,000 – 20,000 chars	Send content, strip structural summary appendix if present	~3-6k
20,000 – 40,000 chars	Send structural summary + first 5,000 chars of body content	~3k
> 40,000 chars	Re-run with `--summary-only`, send summary only	~1-2k

Implementation: When reading content_analysis.json, check each file's total_chars. If content exceeds the threshold, extract only the needed portion before inserting into the LLM prompt. Do NOT blindly paste the entire content field into the prompt.

Step 3.2 — Generate slide plan (LLM)

text

You are a professional presentation designer. Generate a complete slide plan in JSON format.

PRESENTATION PURPOSE: {purpose}

CONTENT SOURCES:
{content from content_analysis.json, or raw_topic if no files}

STYLE REQUIREMENTS:
- Category: {suggested_category}
- Palette: primary={primary_hex}, secondary={secondary_hex}, accent={accent_hex}
- Font heading: {font_heading}, Font body: {font_body}
- Tone: {suggested_tone}
- Target slides: {slide_count}
- Language: Match the language of the content sources

OUTPUT FORMAT — respond ONLY with valid JSON, no markdown fences:
{
  "title": "Presentation Title",
  "style_category": "{suggested_category}",
  "palette": {
    "primary": "HEX_NO_HASH",
    "secondary": "HEX_NO_HASH",
    "accent": "HEX_NO_HASH"
  },
  "font_heading": "Calibri",
  "font_body": "Calibri",
  "slides": [
    {
      "index": 1,
      "layout": "title",
      "title": "Main Title",
      "subtitle": "Subtitle / tagline / date",
      "body": [],
      "notes": "Speaker notes for this slide",
      "chart": null,
      "visual_hint": ""
    },
    {
      "index": 5,
      "layout": "content",
      "title": "Sales Performance 2024",
      "subtitle": "Year-over-year comparison",
      "body": ["Revenue grew 23% YoY", "NPS score reached 87"],
      "notes": "Emphasize the Q4 recovery story",
      "chart": {
        "type": "bar",
        "title": "Quarterly Revenue (万元)",
        "data": [
          {"label": "Q1", "value": 120},
          {"label": "Q2", "value": 145},
          {"label": "Q3", "value": 98},
          {"label": "Q4", "value": 200}
        ]
      },
      "visual_hint": ""
    },
    {
      "index": 8,
      "layout": "content",
      "title": "Team Highlights",
      "body": ["Cross-functional collaboration", "30% headcount growth"],
      "notes": "Show the team photo here",
      "chart": null,
      "visual_hint": "image:Team photo from 2024 annual offsite"
    }
  ]
}

LAYOUT TYPES:
- "title"      → Cover/conclusion, dark background (slide 1 and last slide)
- "section"    → Section divider, dark background (between major topics)
- "content"    → Standard slide: title + optional subtitle + bullets (max 8)
- "two-column" → Comparison split; body split left/right (max 10 total)
- "quote"      → Large pull quote; body[0] is the quote text (max 1)
- "stat"       → 1-3 big numbers; each body item: "Number|Label" e.g. "85%|Satisfaction" (max 3)

CHART FIELD — use for any slide with numeric data:
  chart.type:  "bar" | "line" | "pie" | "doughnut" | "area"
  chart.title: optional string shown above the chart
  chart.data:  array of {"label": string, "value": number}
               values must be plain numbers — no units, no % signs, no commas
               (e.g. use 85 not "85%" or "85万")
  Set chart to null when no chart is needed.
  When chart is present with body bullets, bullets go left and chart fills right.

VISUAL_HINT FIELD — image/icon placeholders only:
  "image:<description>"  → gray dashed placeholder box (photos, KV art, etc.)
  "icon:<name>"          → small colored accent shape
  ""                     → nothing
  DO NOT put chart data in visual_hint. Use the chart field instead.

RULES:
1. Every slide must have a clear, specific title
2. Body points: max 15 words each
3. Use varied layouts — avoid repeating "content" every slide
4. Use "section" dividers between major topic groups
5. First slide: layout "title". Last slide: "title" or "section"
6. NEVER use "#" prefix in hex colors — "1E2761" not "#1E2761"
7. index must be sequential starting from 1
8. Add chart field for slides with numeric data
9. Use visual_hint only for image/icon placeholders
10. Derive all content faithfully from the provided source materials

Save as slide_plan_raw.json.

Step 3.3 — Validate

bash

python3 scripts/validator.py --plan slide_plan_raw.json --out slide_plan.json 2>validation_report.json
cat validation_report.json

Step 3.4 — Generate PPTX

No template:

bash

python3 scripts/generate_ppt.py --plan slide_plan.json --output presentation.pptx

With template:

bash

python3 scripts/generate_ppt.py \
  --plan slide_plan.json \
  --output presentation.pptx \
  --mode from-template \
  --template uploaded_template.pptx \
  --skip-validation

Patch mode — modify specific slides without regenerating the full plan:

bash

python3 scripts/generate_ppt.py \
  --plan slide_plan.json \
  --output presentation.pptx \
  --patch '{"slides": [{"index": 3, "title": "Updated Title", "body": ["New bullet 1", "New bullet 2"]}]}'

Patch merges specified fields into matching slides by index. Unspecified fields and other slides are preserved. Use this when the user requests changes to specific slides — it avoids a full LLM re-generation of the slide plan.

Step 3.5 — QA

bash

python3 -m markitdown presentation.pptx
python3 -m markitdown presentation.pptx | grep -iE "\bx{3,}\b|lorem|ipsum|\bTODO"

# Visual inspection (requires LibreOffice + pdftoppm):
python3 scripts/office/soffice.py --headless --convert-to pdf presentation.pptx
rm -f slide-*.jpg && pdftoppm -jpeg -r 150 presentation.pdf slide
ls -1 "$PWD"/slide-*.jpg

Full Flow at a Glance

text

Phase 0 ─── check_env.py ──────────────────── First install only
               │
Phase 1.1 ─── Ask: purpose & topic type
               │
Phase 1.2 ─── Ask: content / upload files
               │
Phase 1.3 ─── Ask: PPT template?
               │                    │
              YES                   NO
               │                    │
          Ask: template          Phase 2.1
          or content?            suggest_style.py
               │                 present → confirm
          template                    │
               │                      │
          Phase 2.2               confirmed
          analyze_template            │
          present info                │
               └──────────┬──────────┘
                           │
Phase 3 ─── analyze_content.py
               │
           LLM → slide_plan_raw.json
               │
           validator.py
               │
           generate_ppt.py
               │
           QA → output .pptx ✅

Script Reference

Script	Purpose	Key flags
`check_env.py`	Verify all dependencies (first install)	`--fix`, `--json`
`analyze_content.py`	Parse reference docs	`--max-chars`, `--summary-only`
`analyze_template.py`	Extract template design system	—
`suggest_style.py`	Infer style from content	`--lang`, `--slides`
`validator.py`	Sanitize LLM slide plan	`--out`, `--strict`
`generate_ppt.py`	Build PPTX (both modes)	`--mode`, `--template`, `--skip-validation`, `--patch`
`template_builder.py`	Fill template via python-pptx	`--template`, `--plan`, `--output`
`chart_builder.py`	chart dict → PptxGenJS chart code	(imported by generate_ppt.py)

Design Principles

Every data slide needs a chart — use the chart field for any numeric data
Dark title slides, light content slides — "sandwich" structure
Vary layouts — never repeat the same layout 3+ times in a row
Left-align body text, center titles only
Minimum 0.5" margins — never crowd edges
Never use accent lines under titles — use whitespace instead
NEVER # prefix in hex colors — PptxGenJS file corruption
NEVER reuse PptxGenJS option objects — use makeShadow() factory pattern

Platform Notes

This skill works in both OpenClaw (agent framework with external LLM) and AI IDEs (Claude Code, Cursor, Windsurf, etc.).

Script paths

All script references use relative paths (scripts/xxx.py). When running in an AI IDE, resolve them relative to the skill installation directory. Example:

bash

# If skill is installed at /path/to/ppt-generator/
python3 /path/to/ppt-generator/scripts/suggest_style.py --query "AI trends"

Step 3.2 in AI IDEs

In Claude Code, Cursor, and similar environments, the agent itself IS the LLM. Step 3.2's slide plan prompt should be treated as a self-instruction: generate the JSON directly and write it to slide_plan_raw.json. No external API call is needed.

Dependency installation in sandboxed environments

Some AI IDEs run in sandboxed containers. If check_env.py --fix fails due to permission issues, try:

bash

pip install "markitdown[pptx]" pypdf python-pptx --break-system-packages --user

Dependencies

bash

# Python (required)
pip install "markitdown[pptx]" pypdf python-pptx --break-system-packages

# Node.js (required)
npm install -g pptxgenjs

# System (optional — QA visual inspection only)
# Linux:  sudo apt install libreoffice poppler-utils
# macOS:  brew install --cask libreoffice && brew install poppler
# Windows: see check_env.py output for download links