relief-pitcher-ai

星标7

分支0

更新时间2026年3月28日 15:08

A skill that auto-generates presentation materials for talks (from LTs to 30-min talks). From a theme, it creates an outline, generates PowerPoint/ODP slides, writes talk scripts (speaker notes), synthesizes speech, embeds audio, and exports MP4 video — all in one pipeline. Use cases: (1) Generate a full presentation from a talk theme, (2) Create slides from an existing outline, (3) Add talk scripts to slides, (4) Synthesize speech from scripts and embed into slides, (5) Export the presentation as a video. Triggers: LT, lightning talk, presentation, slides, talk script, relief pitcher, speech synthesis, lecture, video, MP4, 代打, プレゼン作成, スライド作成, 登壇, ライトニングトーク, 音声合成, 講演, 動画。

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

hawkymisc

hawkymisc/daida-ai

打开 GitHub 仓库查看创作者相关仓库

下载

Relief Pitcher AI — Auto-Generate Presentation Skill

Overview

Executes the full pipeline from theme input to a finished presentation. Up through Step 6 (slideshow configuration), the output is a complete PPTX. Step 7 can additionally export as an MP4 video, but requires extra external tools (LibreOffice, ffmpeg). Since many users do not need video, after completing Step 6, confirm whether video is needed; if not, skip Step 7.

Help Display

When the user asks "help", "how to use", "show me the workflow", etc., display the following pipeline diagram:

╔══════════════════════════════════════════════════════════════╗
║               Relief Pitcher AI Pipeline                     ║
╠══════════════════════════════════════════════════════════════╣
║                                                              ║
║  Step 1    Theme → Outline (Markdown)                        ║
║     │                                                        ║
║  Step 1.5  Outline → Slide Spec (JSON)                       ║
║     │                                                        ║
║  Step 1.7  Image Generation (SVG / Gemini) ── Optional       ║
║     │                                                        ║
║  Step 2    Slide Spec → PPTX Generation                      ║
║     │         ↕ User can review & edit the PPTX              ║
║  Step 3    Talk Script Update ──────────── Optional           ║
║     │                                                        ║
║  Step 4    Dict → TTS Script → Speech Synthesis              ║
║     │         ↕ User can review & edit pronunciation         ║
║  Step 5    Audio Files → Embed into PPTX                     ║
║     │                                                        ║
║  Step 6    Slideshow Auto-Play Configuration                 ║
║     │                                                        ║
║     ▼                                                        ║
║  Done!     presentation_final.pptx                           ║
║     │                                                        ║
║  Step 7    MP4 Video Export ───────────── Optional            ║
║     │       (with automatic validation)                      ║
║     ▼                                                        ║
║  Video!    presentation.mp4                                  ║
║                                                              ║
╠══════════════════════════════════════════════════════════════╣
║  Templates: tech / casual / formal                           ║
║  TTS: edge (default) / voicevox                              ║
║  Output: PPTX / ODP (conversion option) / MP4 (video option) ║
╚══════════════════════════════════════════════════════════════╝

Additional notes (display alongside the pipeline diagram when showing help):

You can run all steps automatically, or resume from any step
Steps marked with ↕ allow the user to manually edit and go back
Step 1.7 (image generation) and Step 3 (script update) are optional
Step 7 (video export) is also optional. It requires additional tools (LibreOffice, ffmpeg); skip if video is not needed

Prerequisites

On first run, execute the setup:

bash ${CLAUDE_SKILL_DIR}/scripts/setup.sh

setup.sh installs Python dependencies and then automatically checks availability of video generation tools (LibreOffice, ffmpeg, pdftoppm). If any tools show [--] in the output, and the user wants video output, guide them through installation. If video is confirmed unnecessary, the tool-not-found warnings can be ignored.

Common Pattern for Script Execution

All Python scripts are run via the run.sh wrapper (automatic venv resolution):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh <script_name.py> [args...]

Workflow Selection (Step Resume Support)

Confirm with the user:

Full run: Auto-generate everything from a theme
Step-specific: Start from a specific step (using existing files)

Confirm the output directory with the user (default: ./output/)

Video Output Confirmation

At the start of the workflow, confirm whether MP4 video is also needed. If the AskUserQuestion tool is available, always use it to confirm.

After the PPTX is complete, you can also export it as an MP4 video (Step 7). Video export requires LibreOffice and ffmpeg to be installed. Would you like to create a video as well? (We can skip it if not needed.)

"Yes" / "Video too" / "MP4 too" → Execute through Step 7. Check tool availability during setup.sh
"No" / "Not needed" / "PPTX only" → Complete at Step 6. Skip the Step 7 prompt entirely
No answer / Ambiguous → Confirm again at Step 6 completion (see Step 6 guidance section)

Step Resume / Interruption

After the user edits intermediate files (PPTX, TTS scripts, etc.), the workflow can be resumed from a specified step.

When the user says "I want to redo from Step N":

Check that the "Required Input" from the table below is available
If anything is missing, inform the user
If everything is ready, execute sequentially from that step to the end

Resume Step	Required Input	Typical Use Case
Step 1	Theme (verbal)	Start over from scratch
Step 1.5	`output/outline.md`	Manually edited the outline
Step 1.7	`output/slide_spec.json`	Manually edited the JSON spec
Step 2	`output/slide_spec.json`	Manually edited the JSON spec
Step 3	`output/presentation.pptx`	Manually edited the PPTX
Step 4	`output/presentation.pptx`	Changed scripts / want to fix pronunciation
Step 5	`output/presentation.pptx` + `output/audio/`	Replaced audio files
Step 6	`output/presentation_with_audio.pptx`	Edited the audio-embedded PPTX
Step 7	`output/presentation_final.pptx` + `output/audio/`	Want to regenerate the video

Common interruption patterns:

"I edited the PPTX, recreate the audio" → Resume from Step 4
"I fixed the pronunciation, just redo the audio" → Resume from Step 4c (with --script)
"I want to change the slide content" → Resume from Step 1.5 (edit JSON → Step 2 onward)
"I want to change the template" → Change metadata.template in Step 1.5 → Resume from Step 2
"I want to change the TTS engine" → Resume from Step 4 (change --engine)
"I just want to regenerate the video" → Resume from Step 7

Step 1: Outline Generation

What you (Claude) should do

Gather the following from the user:

Theme: What the talk is about
Audience: Who the LT is for
Duration: How many minutes (default: 5 min)
Event name: Optional

Generate a Markdown outline with the following structure:

# Presentation Title

## Introduction
- Point 1
- Point 2

## Main Topic 1: Section Name
- Point
- Point

## Main Topic 2: Section Name
- Point
- Point

## Summary
- Point

Guidelines:

LT (5 min): 3-4 sections, 2-3 items each, 10 slides max
Standard talk (15 min): 5-7 sections, 30 slides max
Lecture (30 min): 7-10 sections, 60 slides max
Target 30 sec to 1 min per slide

Script Execution

Save the generated Markdown to a file, then pass it to the script:

# First save the outline Markdown to a file (using the Write tool)
# Then use the script to save to the designated path
bash ${CLAUDE_SKILL_DIR}/scripts/run.sh generate_outline.py output/outline.md --stdin < output/outline_draft.md

Step 1.5: Content Enrichment

What you (Claude) should do

Read the outline generated in Step 1 and create a slide spec JSON following these guidelines:

Enrichment guidelines:

Add specific numbers and data ("fast" → "10x faster")
Use two_content layout when there are comparisons or contrasts
Follow the one-message-per-slide principle
Limit bullet points to 3-5 items (recommended; guardrails allow 1-8)
For code examples, use title_only + code block in the note
The first slide must be title_slide
Use section_header for section dividers

Confirm with the user:

Template: tech (default) / casual / formal
Diagram density: low / medium (default) / high

Diagram Density Levels

Adjust the proportion of slides with an image field according to the density level.

Density	% of slides with image	Guideline
Low	~20% (1-2 out of 10)	Key diagrams only. Primarily text and bullet points
Medium	~40% (3-4 out of 10)	One diagram per section. Moderate use of flowcharts and comparison charts
High	~60% (5-6 out of 10)	Visual elements on nearly every slide. Actively use architecture diagrams, charts, and illustrations

Default density by template (determined in Step 1.5):

Template	Default Density
`tech`	Medium
`casual`	Medium
`formal`	Low

If the user explicitly specifies a density, use that. Otherwise, use the template default.

Reference: Talk style and recommended density (selected in Step 3):

Talk Style	Recommended Density
`casual`	Medium
`keynote`	High
`formal`	Low
`humorous`	Medium

Talk style is selected in Step 3, so at the Step 1.5 stage, use the template default. Templates and talk styles that share a name (casual, formal) correspond to the same density.

JSON Format

{
  "metadata": {
    "title": "Presentation Title",
    "subtitle": "Speaker Name",
    "event": "Event Name",
    "template": "tech"
  },
  "slides": [
    {
      "layout": "title_slide",
      "title": "Presentation Title",
      "subtitle": "2026/03/10 @ Event Name - Speaker Name",
      "note": "Hello everyone. Today I'll be talking about X."
    },
    {
      "layout": "section_header",
      "title": "Section Name",
      "note": "Let's start by looking at X."
    },
    {
      "layout": "title_and_content",
      "title": "Slide Title",
      "body": ["Point 1", "Point 2", "Point 3"],
      "note": "Talk script (speaker notes)"
    },
    {
      "layout": "two_content",
      "title": "Comparison Title",
      "left": {"heading": "Left Heading", "body": ["Item 1", "Item 2"]},
      "right": {"heading": "Right Heading", "body": ["Item 1", "Item 2"]},
      "note": "Talk script"
    },
    {
      "layout": "title_only",
      "title": "Diagram / Code Slide",
      "note": "Explanation for this slide"
    },
    {
      "layout": "title_only",
      "title": "Architecture Diagram",
      "image": "images/architecture.png",
      "note": "This diagram shows the overall system architecture"
    },
    {
      "layout": "blank",
      "image": "images/fullscreen_photo.jpg",
      "note": "This photo shows the project results"
    }
  ]
}

Available layouts: title_slide, section_header, title_and_content, two_content, title_only, blank

Image insertion: Specify an image file path in the image field to insert an image into the slide.

Paths are relative to the slide spec JSON file (e.g., if the spec is output/slide_spec.json, then images/foo.png resolves to output/images/foo.png)
Images are automatically fitted within the content area while maintaining aspect ratio, centered both horizontally and vertically
Images are placed behind text placeholders
Available on all layouts. Recommended: title_only (title + diagram), blank (fullscreen diagram)
Supported formats: PNG, JPEG, GIF, BMP, TIFF, SVG (auto-converted to PNG; requires cairosvg)

Important: All slides must include a talk script (speaker notes) in the note field (including title_slide and section_header). Slides without notes will have no audio generated and will advance quickly (3 seconds) during the slideshow. Notes can be updated in Step 3, but generate a first draft here.

Script Execution

Save the generated JSON to a file, then pass it to the script (piping via echo may fail for large JSON):

# First save the slide spec JSON to a file (using the Write tool)
# Then run the script for validation & saving (specify duration with --duration)
bash ${CLAUDE_SKILL_DIR}/scripts/run.sh enrich_outline.py output/slide_spec.json --stdin --duration 5 < output/slide_spec_draft.json

For a 15-minute talk:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh enrich_outline.py output/slide_spec.json --stdin --duration 15 < output/slide_spec_draft.json

--duration automatically adjusts the slide count limit and estimated speaking time limit.

Step 1.7: Image Generation (Optional)

Generate images when you want to include diagrams or illustrations in slides.

Choosing a Method

Follow this flowchart to decide the method:

1. Is GEMINI_API_KEY set?
   ├─ No → Generate everything as SVG (see "SVG Generation" section below)
   └─ Yes → Go to 2
2. What is the image content?
   ├─ SVG-suited (see list below) → Go to "SVG Generation" section
   └─ Nano Banana-suited (see list below) → Go to "Nano Banana" section

SVG-suited (Claude generates directly)

Use SVG for images matching the following. No API key needed, instant generation, editable.

Flowcharts, process diagrams, step diagrams
Architecture diagrams, system configuration diagrams
Comparison charts (Before/After, A vs B)
Bar charts, pie charts, simple charts
Icons, logos, symbols
Timelines, roadmaps
Tables, matrix diagrams
Text-heavy concept diagrams, mind maps

Nano Banana-suited (generated via Gemini API)

Use Nano Banana for images matching the following. Requires GEMINI_API_KEY.

Photo-like backgrounds, landscapes, people
Realistic texture illustrations, 3D rendering style
Specific object depictions (product photo style, etc.)
Hand-drawn, watercolor, oil painting style art
Screenshot-style mockups

When in doubt, use SVG. SVG can be generated instantly and is easy to edit. Use Nano Banana only when you need "realism that SVG cannot express."

Confirm with the user

Which slides need images, and the content for each
For Nano Banana: aspect ratio (16:9 / 4:3 / 1:1), resolution (1K / 2K / 4K)

Nano Banana (Gemini Image Generation API)

Prerequisites: GEMINI_API_KEY environment variable

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh generate_image.py \
  --prompt "A professional conference stage with spotlight, modern tech event atmosphere" \
  --output output/images/slide1_background.png \
  --aspect-ratio 16:9 \
  --size 1K \
  --model pro

Alias	Model ID	Use Case
`pro`	`gemini-3-pro-image-preview`	High quality, complex prompts, text rendering
`flash`	`gemini-3.1-flash-image-preview`	Fast generation, bulk generation
`legacy`	`gemini-2.5-flash-image`	Legacy model

After generation, show the image to the user with the Read tool and adjust the prompt as needed.

SVG Generation (No API Required)

You (Claude) generate SVG code directly. SVG files can be specified directly in the slide spec JSON image field (automatically converted to PNG at build time).

Prerequisites: cairosvg must be installed (auto-installed if setup.sh has been run). If not installed:

pip install cairosvg

Procedure

Create the SVG file in output/images/ using the Write tool
Specify the SVG path directly in the slide spec JSON image field
It will be automatically converted to PNG and inserted when create_slides.py runs in Step 2

Manual conversion is not needed. However, if you want to preview beforehand:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh svg_to_png.py input.svg output.png

SVG Sizes

Use	viewBox	Description
Full slide	`0 0 1920 1080`	16:9, suited for blank layout
Content area	`0 0 1200 900`	4:3, suited for title_only layout
Icon / Logo	`0 0 400 400`	1:1

SVG Font Size Requirements

SVG font-size is in px units within the viewBox coordinate system, but when embedded in PPTX, the image is scaled down, so the actual display size is determined by:

rendered_pt = f_svg × display_w_emu / (viewBox_w × 12700)

display_w_emu = min(max_w, max_h × viewBox_w / viewBox_h)
max_w = slide_w − 2 × 457200 (left/right margins)
max_h = slide_h − img_top − 457200 (title_only: img_top=1600200, blank: img_top=457200)
12700 = EMU per 1pt (OOXML standard)

Minimum standard for presentation materials: 12pt. Follow these minimum SVG font-size values:

viewBox	Layout	Min font-size	Recommended body	Recommended heading
`1920×1080`	title_only	35 px	40 px	56 px
`1920×1080`	blank	28 px	32 px	48 px
`1200×900`	title_only	29 px	32 px	48 px
`1200×900`	blank	24 px	28 px	40 px

Note: Font sizes in SVGs are automatically validated when create_slides.py runs. A warning will be output if any text falls below the minimum.

Color Scheme by Template

Template	Background	Accent 1	Accent 2	Text
tech	`#1E293B`	`#38BDF8`	`#818CF8`	`#E2E8F0`
casual	`#FFF8F0`	`#FF6B35`	`#06D6A0`	`#2D3748`
formal	`#FFFFFF`	`#1B2D45`	`#C49B66`	`#1B2D45`

Note: Always specify font-family="sans-serif" for fonts. cairosvg automatically injects a Japanese font fallback, so Japanese text will not display as tofu (□). If font-family already contains Hiragino / Yu Gothic / Noto Sans CJK JP, injection is skipped.

SVG Pattern Examples

Flowchart (3 steps):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <defs>
    <marker id="arr" markerWidth="10" markerHeight="7" refX="10" refY="3.5" orient="auto">
      <polygon points="0 0,10 3.5,0 7" fill="#818CF8"/>
    </marker>
  </defs>
  <!-- Step 1 -->
  <rect x="160" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="360" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 1</text>
  <!-- Arrow 1→2 -->
  <line x1="560" y1="500" x2="720" y2="500" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Step 2 -->
  <rect x="760" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="960" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 2</text>
  <!-- Arrow 2→3 -->
  <line x1="1160" y1="500" x2="1320" y2="500" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Step 3 -->
  <rect x="1360" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="1560" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 3</text>
</svg>

Comparison Chart (Before / After):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <!-- Before -->
  <rect x="80" y="120" width="840" height="840" rx="20" fill="#334155" stroke="#475569" stroke-width="2"/>
  <text x="500" y="200" text-anchor="middle" fill="#EF4444"
        font-size="56" font-family="sans-serif" font-weight="bold">Before</text>
  <text x="500" y="400" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">3 hours manual</text>
  <text x="500" y="460" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">Error-prone</text>
  <text x="500" y="520" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">Person-dependent</text>
  <!-- After -->
  <rect x="1000" y="120" width="840" height="840" rx="20" fill="#334155" stroke="#38BDF8" stroke-width="2"/>
  <text x="1420" y="200" text-anchor="middle" fill="#38BDF8"
        font-size="56" font-family="sans-serif" font-weight="bold">After</text>
  <text x="1420" y="400" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">5 min automated</text>
  <text x="1420" y="460" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">Consistent quality</text>
  <text x="1420" y="520" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">Anyone can run it</text>
</svg>

Architecture Diagram (3 layers):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <defs>
    <marker id="arr" markerWidth="10" markerHeight="7" refX="10" refY="3.5" orient="auto">
      <polygon points="0 0,10 3.5,0 7" fill="#818CF8"/>
    </marker>
  </defs>
  <!-- Frontend -->
  <rect x="660" y="80" width="600" height="140" rx="16" fill="#38BDF8"/>
  <text x="960" y="160" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Frontend</text>
  <!-- Arrow -->
  <line x1="960" y1="220" x2="960" y2="360" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- API -->
  <rect x="660" y="380" width="600" height="140" rx="16" fill="#818CF8"/>
  <text x="960" y="460" text-anchor="middle" fill="#FFFFFF"
        font-size="40" font-family="sans-serif" font-weight="bold">API Server</text>
  <!-- Arrow -->
  <line x1="960" y1="520" x2="960" y2="660" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Database -->
  <rect x="660" y="680" width="600" height="140" rx="16" fill="#334155" stroke="#38BDF8" stroke-width="2"/>
  <text x="960" y="760" text-anchor="middle" fill="#38BDF8"
        font-size="40" font-family="sans-serif" font-weight="bold">Database</text>
</svg>

Workflow

In the Step 1.5 slide spec JSON, specify the output path in the image field
- Nano Banana: specify a .png path
- SVG: specify a .svg path (auto-converted)
Generate images in this step
Images are automatically inserted when slides are generated in Step 2

Step 2: Slide Creation

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh create_slides.py output/slide_spec.json output/presentation.pptx

To use a custom template:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh create_slides.py output/slide_spec.json output/presentation.pptx --template path/to/template.pptx

After generation, ask the user to review the PPTX.

Step 3: Talk Script Update (Optional)

If notes are already included from Step 1.5, this step can be skipped. Execute when the user wants to change the writing style or regenerate scripts.

What you (Claude) should do

Read the current notes:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh write_talk_script.py output/presentation.pptx --read

Refer to references/talk-styles.md and regenerate scripts for all slides in the user's preferred style
- Styles: casual (default) / keynote / formal / humorous
Save as a JSON array and write to the PPTX:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh write_talk_script.py output/presentation.pptx --notes-json output/notes.json --output output/presentation.pptx

Step 4: Speech Synthesis

Confirm with the user

TTS engine: edge (default) / voicevox
Voice: default is ja-JP-NanamiNeural (edge), 1 = Zundamon (voicevox)

Step 4a: TTS Script Export

Export speaker notes as a text file for text-to-speech. When a pronunciation dictionary (--dict) is specified, common mispronunciation patterns are automatically replaced.

With pronunciation dictionary (recommended):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh export_tts_script.py output/presentation.pptx output/tts_script.txt --dict ${CLAUDE_SKILL_DIR}/assets/pronunciation_dict.tsv

Without pronunciation dictionary:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh export_tts_script.py output/presentation.pptx output/tts_script.txt

Step 4b: Review and Edit Pronunciation Text

Present the exported script file contents to the user. Prompt them to check for items not auto-corrected by the dictionary, or project-specific terminology.

Examples auto-corrected by dictionary:

"生成" → "せいせい", "Claude" → "クロード", "LLM" → "エルエルエム"

Cases requiring manual correction:

Project-specific abbreviations and product names
Technical terms not registered in the dictionary

If the user decides no corrections are needed, proceed to Step 4c (synthesize without --script, as before). If entries should be added to the dictionary, update pronunciation_dict.tsv.

Step 4c: Script Execution

With a corrected script file (--script):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine edge --script output/tts_script.txt

Without corrections (synthesize directly from PPTX notes, as before):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine edge

When using VOICEVOX (VOICEVOX Engine must be running beforehand):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine voicevox --script output/tts_script.txt

TTS Failure Handling

If the TTS API fails for some or all slides, failed slides are skipped and audio files are generated only for successful ones (partial success).

The script reports failure count to stderr and exits normally
Slides without audio will auto-advance with default timing (3 seconds) in Step 6
After TTS recovery, re-running Step 4c will generate audio for all slides
The PPTX from Step 2 remains usable as-is (presentation is possible without audio)

Step 5: Audio Embedding

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh embed_audio.py output/presentation.pptx output/audio/ output/presentation_with_audio.pptx

Step 6: Slideshow Auto-Play Configuration

Configures auto-advance and audio auto-play on the audio-embedded PPTX. In PowerPoint (Windows / macOS), starting the slideshow plays it fully automatically to the end. In LibreOffice Impress, auto-advance does not work, so slides must be advanced manually (see "Cross-Platform Compatibility Notes" for details).

Slides with audio: auto-advance after audio playback + buffer (default 1 second)
Slides without audio (title, section headers, etc.): auto-advance with fixed time (default 3 seconds)

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_slideshow.py output/presentation_with_audio.pptx output/presentation_final.pptx

To adjust display timing:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_slideshow.py output/presentation_with_audio.pptx output/presentation_final.pptx --silent-duration 5000 --audio-buffer 2000

User Guidance (Required)

After completing the slideshow configuration, always communicate the following notes to the user:

Pre-Slideshow Checklist

For PowerPoint (Windows / macOS): Verify that "Use Timings" is checked in the slideshow settings.

Windows: "Slide Show" tab → "Set Up Slide Show" → Check "Use Timings"

macOS: "Slide Show" menu → "Set Up Show..." → "Options" → Check "Use Timings"

If this is unchecked, auto-advance and audio auto-play will not work.

For LibreOffice Impress: Auto-advance does not work. Please advance slides manually during the slideshow (click or arrow keys). Audio will auto-play on each slide.

Step 7 Guidance (Required)

Following the Step 6 guidance, always confirm the following:

This presentation can also be exported as an MP4 video. This is useful for posting to YouTube or sharing in environments where PowerPoint is not available. However, additional tools (LibreOffice, ffmpeg) need to be installed. If you don't need video, the process is complete here. Would you like to create a video as well?

If the user responds with "Not needed", "No thanks", etc., skip Step 7 and end the pipeline.

Step 7: MP4 Video Export (Optional)

Generate an MP4 video from the PPTX and audio files. Execute when you want to distribute or share the slideshow as a video.

Confirm with the user

When the user responds "Create the video" to the Step 7 guidance, confirm the following:

Video frame rate (default: 30fps)
Display duration for slides without audio (default: 3 seconds)

Prerequisites

Video generation requires additional tools. Availability is shown during setup.sh execution.

Tool	Purpose	Required
LibreOffice	PPTX → PDF → PNG rendering	Yes
ffmpeg	Video clip generation & concatenation	Yes
pdftoppm (poppler-utils)	PDF → PNG conversion (high quality)	Recommended (ffmpeg can substitute)

If not installed:

# Ubuntu/Debian
sudo apt install libreoffice ffmpeg poppler-utils

# macOS
brew install libreoffice ffmpeg poppler

Note on snap version of LibreOffice: Use an output directory under $HOME. Writing to /tmp etc. may fail due to sandbox restrictions.

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_video.py output/presentation_final.pptx output/audio output/presentation.mp4

With options:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_video.py output/presentation_final.pptx output/audio output/presentation.mp4 --fps 30 --silent-duration 3.0

Automatic Validation

After generation, the following checks are automatically performed:

Check Item	Details
File existence	MP4 file exists and size > 0
ffprobe read	File is not corrupted
Codecs	Video: H.264 / Audio: AAC
Resolution	Width and height are even numbers (libx264 compatible)
Duration	Positive value

If validation fails, error details are displayed. Use --skip-validation to disable.

User Guidance

After video generation is complete, communicate the following:

About the Video Output

presentation.mp4 has been generated

Display time for each slide is based on the actual audio file duration

Slides without audio display for 3 seconds by default

Playable with common video players (VLC, QuickTime, etc.)

Format Conversion (Optional)

If ODP format is needed:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh convert_format.py output/presentation.pptx --outdir output/

Prerequisite: LibreOffice must be installed.

Cross-Platform Compatibility Notes

macOS PowerPoint Compatibility

PPTX files generated by Relief Pitcher AI work on both macOS and Windows PowerPoint as well as LibreOffice Impress. Note the following technical constraints.

Audio Auto-Play (Step 6)

Method	macOS	Windows	LibreOffice
`p:cmd type="call" cmd="playFrom(0)"`	Yes	Yes	Yes
`p:audio > p:cMediaNode`	No (treated as corrupt)	Yes	Yes

macOS PowerPoint treats p:audio > p:cMediaNode (OOXML media node method) as a corrupted file
Instead, use p:cmd type="call" cmd="playFrom(0)" (command animation method)
Specify nodeType="afterEffect" + grpId="0" for "auto-execute after previous animation"
This method also works correctly on Windows PowerPoint / LibreOffice Impress

Audio Icon (Step 5)

Icon	macOS	Windows
32x32 visible icon	Yes, displayed	Yes, displayed
1x1 transparent PNG	No, invisible	Yes, auto-replaced with speaker icon

macOS PowerPoint displays transparent icons as-is (Windows auto-replaces with a speaker icon)
Therefore, a 32x32 visible speaker icon is used when embedding audio
The icon is dynamically positioned at the bottom-right corner based on slide dimensions (0.25" margin from right and bottom edges)
Correctly positioned at bottom-right for both 4:3 and 16:9 aspect ratios

Auto-Advance

Method	macOS	Windows	LibreOffice
`p:transition advTm`	Yes	Yes	No (not supported)

LibreOffice Impress does not support auto-advance. Even with timing settings (advTm, mainSeq.dur) correctly configured in the PPTX, LibreOffice ignores them — this is a known limitation (Bug 101527)
To play in LibreOffice, advance slides manually or manually set "auto-transition" time for each slide in LibreOffice's "Slide Transition" panel
Auto-advance works correctly in PowerPoint (Windows / macOS)

Reference

references/pptx-guide.md: Details on layouts and placeholders
references/talk-styles.md: Talk script style definitions
references/tts-plugins.md: How to add TTS plugins

name

relief-pitcher-ai

description

Relief Pitcher AI — Auto-Generate Presentation Skill

Overview

Help Display

When the user asks "help", "how to use", "show me the workflow", etc., display the following pipeline diagram:

╔══════════════════════════════════════════════════════════════╗
║               Relief Pitcher AI Pipeline                     ║
╠══════════════════════════════════════════════════════════════╣
║                                                              ║
║  Step 1    Theme → Outline (Markdown)                        ║
║     │                                                        ║
║  Step 1.5  Outline → Slide Spec (JSON)                       ║
║     │                                                        ║
║  Step 1.7  Image Generation (SVG / Gemini) ── Optional       ║
║     │                                                        ║
║  Step 2    Slide Spec → PPTX Generation                      ║
║     │         ↕ User can review & edit the PPTX              ║
║  Step 3    Talk Script Update ──────────── Optional           ║
║     │                                                        ║
║  Step 4    Dict → TTS Script → Speech Synthesis              ║
║     │         ↕ User can review & edit pronunciation         ║
║  Step 5    Audio Files → Embed into PPTX                     ║
║     │                                                        ║
║  Step 6    Slideshow Auto-Play Configuration                 ║
║     │                                                        ║
║     ▼                                                        ║
║  Done!     presentation_final.pptx                           ║
║     │                                                        ║
║  Step 7    MP4 Video Export ───────────── Optional            ║
║     │       (with automatic validation)                      ║
║     ▼                                                        ║
║  Video!    presentation.mp4                                  ║
║                                                              ║
╠══════════════════════════════════════════════════════════════╣
║  Templates: tech / casual / formal                           ║
║  TTS: edge (default) / voicevox                              ║
║  Output: PPTX / ODP (conversion option) / MP4 (video option) ║
╚══════════════════════════════════════════════════════════════╝

Additional notes (display alongside the pipeline diagram when showing help):

You can run all steps automatically, or resume from any step
Steps marked with ↕ allow the user to manually edit and go back
Step 1.7 (image generation) and Step 3 (script update) are optional
Step 7 (video export) is also optional. It requires additional tools (LibreOffice, ffmpeg); skip if video is not needed

Prerequisites

On first run, execute the setup:

bash ${CLAUDE_SKILL_DIR}/scripts/setup.sh

Common Pattern for Script Execution

All Python scripts are run via the run.sh wrapper (automatic venv resolution):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh <script_name.py> [args...]

Workflow Selection (Step Resume Support)

Confirm with the user:

Full run: Auto-generate everything from a theme
Step-specific: Start from a specific step (using existing files)

Confirm the output directory with the user (default: ./output/)

Video Output Confirmation

At the start of the workflow, confirm whether MP4 video is also needed. If the AskUserQuestion tool is available, always use it to confirm.

After the PPTX is complete, you can also export it as an MP4 video (Step 7). Video export requires LibreOffice and ffmpeg to be installed. Would you like to create a video as well? (We can skip it if not needed.)

"Yes" / "Video too" / "MP4 too" → Execute through Step 7. Check tool availability during setup.sh
"No" / "Not needed" / "PPTX only" → Complete at Step 6. Skip the Step 7 prompt entirely
No answer / Ambiguous → Confirm again at Step 6 completion (see Step 6 guidance section)

Step Resume / Interruption

After the user edits intermediate files (PPTX, TTS scripts, etc.), the workflow can be resumed from a specified step.

When the user says "I want to redo from Step N":

Check that the "Required Input" from the table below is available
If anything is missing, inform the user
If everything is ready, execute sequentially from that step to the end

Resume Step	Required Input	Typical Use Case
Step 1	Theme (verbal)	Start over from scratch
Step 1.5	`output/outline.md`	Manually edited the outline
Step 1.7	`output/slide_spec.json`	Manually edited the JSON spec
Step 2	`output/slide_spec.json`	Manually edited the JSON spec
Step 3	`output/presentation.pptx`	Manually edited the PPTX
Step 4	`output/presentation.pptx`	Changed scripts / want to fix pronunciation
Step 5	`output/presentation.pptx` + `output/audio/`	Replaced audio files
Step 6	`output/presentation_with_audio.pptx`	Edited the audio-embedded PPTX
Step 7	`output/presentation_final.pptx` + `output/audio/`	Want to regenerate the video

Common interruption patterns:

"I edited the PPTX, recreate the audio" → Resume from Step 4
"I fixed the pronunciation, just redo the audio" → Resume from Step 4c (with --script)
"I want to change the slide content" → Resume from Step 1.5 (edit JSON → Step 2 onward)
"I want to change the template" → Change metadata.template in Step 1.5 → Resume from Step 2
"I want to change the TTS engine" → Resume from Step 4 (change --engine)
"I just want to regenerate the video" → Resume from Step 7

Step 1: Outline Generation

What you (Claude) should do

Gather the following from the user:

Theme: What the talk is about
Audience: Who the LT is for
Duration: How many minutes (default: 5 min)
Event name: Optional

Generate a Markdown outline with the following structure:

# Presentation Title

## Introduction
- Point 1
- Point 2

## Main Topic 1: Section Name
- Point
- Point

## Main Topic 2: Section Name
- Point
- Point

## Summary
- Point

Guidelines:

LT (5 min): 3-4 sections, 2-3 items each, 10 slides max
Standard talk (15 min): 5-7 sections, 30 slides max
Lecture (30 min): 7-10 sections, 60 slides max
Target 30 sec to 1 min per slide

Script Execution

Save the generated Markdown to a file, then pass it to the script:

# First save the outline Markdown to a file (using the Write tool)
# Then use the script to save to the designated path
bash ${CLAUDE_SKILL_DIR}/scripts/run.sh generate_outline.py output/outline.md --stdin < output/outline_draft.md

Step 1.5: Content Enrichment

What you (Claude) should do

Read the outline generated in Step 1 and create a slide spec JSON following these guidelines:

Enrichment guidelines:

Add specific numbers and data ("fast" → "10x faster")
Use two_content layout when there are comparisons or contrasts
Follow the one-message-per-slide principle
Limit bullet points to 3-5 items (recommended; guardrails allow 1-8)
For code examples, use title_only + code block in the note
The first slide must be title_slide
Use section_header for section dividers

Confirm with the user:

Template: tech (default) / casual / formal
Diagram density: low / medium (default) / high

Diagram Density Levels

Adjust the proportion of slides with an image field according to the density level.

Density	% of slides with image	Guideline
Low	~20% (1-2 out of 10)	Key diagrams only. Primarily text and bullet points
Medium	~40% (3-4 out of 10)	One diagram per section. Moderate use of flowcharts and comparison charts
High	~60% (5-6 out of 10)	Visual elements on nearly every slide. Actively use architecture diagrams, charts, and illustrations

Default density by template (determined in Step 1.5):

Template	Default Density
`tech`	Medium
`casual`	Medium
`formal`	Low

If the user explicitly specifies a density, use that. Otherwise, use the template default.

Reference: Talk style and recommended density (selected in Step 3):

Talk Style	Recommended Density
`casual`	Medium
`keynote`	High
`formal`	Low
`humorous`	Medium

Talk style is selected in Step 3, so at the Step 1.5 stage, use the template default. Templates and talk styles that share a name (casual, formal) correspond to the same density.

JSON Format

{
  "metadata": {
    "title": "Presentation Title",
    "subtitle": "Speaker Name",
    "event": "Event Name",
    "template": "tech"
  },
  "slides": [
    {
      "layout": "title_slide",
      "title": "Presentation Title",
      "subtitle": "2026/03/10 @ Event Name - Speaker Name",
      "note": "Hello everyone. Today I'll be talking about X."
    },
    {
      "layout": "section_header",
      "title": "Section Name",
      "note": "Let's start by looking at X."
    },
    {
      "layout": "title_and_content",
      "title": "Slide Title",
      "body": ["Point 1", "Point 2", "Point 3"],
      "note": "Talk script (speaker notes)"
    },
    {
      "layout": "two_content",
      "title": "Comparison Title",
      "left": {"heading": "Left Heading", "body": ["Item 1", "Item 2"]},
      "right": {"heading": "Right Heading", "body": ["Item 1", "Item 2"]},
      "note": "Talk script"
    },
    {
      "layout": "title_only",
      "title": "Diagram / Code Slide",
      "note": "Explanation for this slide"
    },
    {
      "layout": "title_only",
      "title": "Architecture Diagram",
      "image": "images/architecture.png",
      "note": "This diagram shows the overall system architecture"
    },
    {
      "layout": "blank",
      "image": "images/fullscreen_photo.jpg",
      "note": "This photo shows the project results"
    }
  ]
}

Available layouts: title_slide, section_header, title_and_content, two_content, title_only, blank

Image insertion: Specify an image file path in the image field to insert an image into the slide.

Paths are relative to the slide spec JSON file (e.g., if the spec is output/slide_spec.json, then images/foo.png resolves to output/images/foo.png)
Images are automatically fitted within the content area while maintaining aspect ratio, centered both horizontally and vertically
Images are placed behind text placeholders
Available on all layouts. Recommended: title_only (title + diagram), blank (fullscreen diagram)
Supported formats: PNG, JPEG, GIF, BMP, TIFF, SVG (auto-converted to PNG; requires cairosvg)

Script Execution

Save the generated JSON to a file, then pass it to the script (piping via echo may fail for large JSON):

# First save the slide spec JSON to a file (using the Write tool)
# Then run the script for validation & saving (specify duration with --duration)
bash ${CLAUDE_SKILL_DIR}/scripts/run.sh enrich_outline.py output/slide_spec.json --stdin --duration 5 < output/slide_spec_draft.json

For a 15-minute talk:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh enrich_outline.py output/slide_spec.json --stdin --duration 15 < output/slide_spec_draft.json

--duration automatically adjusts the slide count limit and estimated speaking time limit.

Step 1.7: Image Generation (Optional)

Generate images when you want to include diagrams or illustrations in slides.

Choosing a Method

Follow this flowchart to decide the method:

1. Is GEMINI_API_KEY set?
   ├─ No → Generate everything as SVG (see "SVG Generation" section below)
   └─ Yes → Go to 2
2. What is the image content?
   ├─ SVG-suited (see list below) → Go to "SVG Generation" section
   └─ Nano Banana-suited (see list below) → Go to "Nano Banana" section

SVG-suited (Claude generates directly)

Use SVG for images matching the following. No API key needed, instant generation, editable.

Flowcharts, process diagrams, step diagrams
Architecture diagrams, system configuration diagrams
Comparison charts (Before/After, A vs B)
Bar charts, pie charts, simple charts
Icons, logos, symbols
Timelines, roadmaps
Tables, matrix diagrams
Text-heavy concept diagrams, mind maps

Nano Banana-suited (generated via Gemini API)

Use Nano Banana for images matching the following. Requires GEMINI_API_KEY.

Photo-like backgrounds, landscapes, people
Realistic texture illustrations, 3D rendering style
Specific object depictions (product photo style, etc.)
Hand-drawn, watercolor, oil painting style art
Screenshot-style mockups

When in doubt, use SVG. SVG can be generated instantly and is easy to edit. Use Nano Banana only when you need "realism that SVG cannot express."

Confirm with the user

Which slides need images, and the content for each
For Nano Banana: aspect ratio (16:9 / 4:3 / 1:1), resolution (1K / 2K / 4K)

Nano Banana (Gemini Image Generation API)

Prerequisites: GEMINI_API_KEY environment variable

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh generate_image.py \
  --prompt "A professional conference stage with spotlight, modern tech event atmosphere" \
  --output output/images/slide1_background.png \
  --aspect-ratio 16:9 \
  --size 1K \
  --model pro

Alias	Model ID	Use Case
`pro`	`gemini-3-pro-image-preview`	High quality, complex prompts, text rendering
`flash`	`gemini-3.1-flash-image-preview`	Fast generation, bulk generation
`legacy`	`gemini-2.5-flash-image`	Legacy model

After generation, show the image to the user with the Read tool and adjust the prompt as needed.

SVG Generation (No API Required)

You (Claude) generate SVG code directly. SVG files can be specified directly in the slide spec JSON image field (automatically converted to PNG at build time).

Prerequisites: cairosvg must be installed (auto-installed if setup.sh has been run). If not installed:

pip install cairosvg

Procedure

Create the SVG file in output/images/ using the Write tool
Specify the SVG path directly in the slide spec JSON image field
It will be automatically converted to PNG and inserted when create_slides.py runs in Step 2

Manual conversion is not needed. However, if you want to preview beforehand:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh svg_to_png.py input.svg output.png

SVG Sizes

Use	viewBox	Description
Full slide	`0 0 1920 1080`	16:9, suited for blank layout
Content area	`0 0 1200 900`	4:3, suited for title_only layout
Icon / Logo	`0 0 400 400`	1:1

SVG Font Size Requirements

SVG font-size is in px units within the viewBox coordinate system, but when embedded in PPTX, the image is scaled down, so the actual display size is determined by:

rendered_pt = f_svg × display_w_emu / (viewBox_w × 12700)

display_w_emu = min(max_w, max_h × viewBox_w / viewBox_h)
max_w = slide_w − 2 × 457200 (left/right margins)
max_h = slide_h − img_top − 457200 (title_only: img_top=1600200, blank: img_top=457200)
12700 = EMU per 1pt (OOXML standard)

Minimum standard for presentation materials: 12pt. Follow these minimum SVG font-size values:

viewBox	Layout	Min font-size	Recommended body	Recommended heading
`1920×1080`	title_only	35 px	40 px	56 px
`1920×1080`	blank	28 px	32 px	48 px
`1200×900`	title_only	29 px	32 px	48 px
`1200×900`	blank	24 px	28 px	40 px

Note: Font sizes in SVGs are automatically validated when create_slides.py runs. A warning will be output if any text falls below the minimum.

Color Scheme by Template

Template	Background	Accent 1	Accent 2	Text
tech	`#1E293B`	`#38BDF8`	`#818CF8`	`#E2E8F0`
casual	`#FFF8F0`	`#FF6B35`	`#06D6A0`	`#2D3748`
formal	`#FFFFFF`	`#1B2D45`	`#C49B66`	`#1B2D45`

SVG Pattern Examples

Flowchart (3 steps):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <defs>
    <marker id="arr" markerWidth="10" markerHeight="7" refX="10" refY="3.5" orient="auto">
      <polygon points="0 0,10 3.5,0 7" fill="#818CF8"/>
    </marker>
  </defs>
  <!-- Step 1 -->
  <rect x="160" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="360" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 1</text>
  <!-- Arrow 1→2 -->
  <line x1="560" y1="500" x2="720" y2="500" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Step 2 -->
  <rect x="760" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="960" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 2</text>
  <!-- Arrow 2→3 -->
  <line x1="1160" y1="500" x2="1320" y2="500" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Step 3 -->
  <rect x="1360" y="420" width="400" height="160" rx="16" fill="#38BDF8"/>
  <text x="1560" y="510" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Step 3</text>
</svg>

Comparison Chart (Before / After):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <!-- Before -->
  <rect x="80" y="120" width="840" height="840" rx="20" fill="#334155" stroke="#475569" stroke-width="2"/>
  <text x="500" y="200" text-anchor="middle" fill="#EF4444"
        font-size="56" font-family="sans-serif" font-weight="bold">Before</text>
  <text x="500" y="400" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">3 hours manual</text>
  <text x="500" y="460" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">Error-prone</text>
  <text x="500" y="520" text-anchor="middle" fill="#94A3B8"
        font-size="40" font-family="sans-serif">Person-dependent</text>
  <!-- After -->
  <rect x="1000" y="120" width="840" height="840" rx="20" fill="#334155" stroke="#38BDF8" stroke-width="2"/>
  <text x="1420" y="200" text-anchor="middle" fill="#38BDF8"
        font-size="56" font-family="sans-serif" font-weight="bold">After</text>
  <text x="1420" y="400" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">5 min automated</text>
  <text x="1420" y="460" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">Consistent quality</text>
  <text x="1420" y="520" text-anchor="middle" fill="#E2E8F0"
        font-size="40" font-family="sans-serif">Anyone can run it</text>
</svg>

Architecture Diagram (3 layers):

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080">
  <rect width="1920" height="1080" fill="#1E293B"/>
  <defs>
    <marker id="arr" markerWidth="10" markerHeight="7" refX="10" refY="3.5" orient="auto">
      <polygon points="0 0,10 3.5,0 7" fill="#818CF8"/>
    </marker>
  </defs>
  <!-- Frontend -->
  <rect x="660" y="80" width="600" height="140" rx="16" fill="#38BDF8"/>
  <text x="960" y="160" text-anchor="middle" fill="#1E293B"
        font-size="40" font-family="sans-serif" font-weight="bold">Frontend</text>
  <!-- Arrow -->
  <line x1="960" y1="220" x2="960" y2="360" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- API -->
  <rect x="660" y="380" width="600" height="140" rx="16" fill="#818CF8"/>
  <text x="960" y="460" text-anchor="middle" fill="#FFFFFF"
        font-size="40" font-family="sans-serif" font-weight="bold">API Server</text>
  <!-- Arrow -->
  <line x1="960" y1="520" x2="960" y2="660" stroke="#818CF8" stroke-width="4" marker-end="url(#arr)"/>
  <!-- Database -->
  <rect x="660" y="680" width="600" height="140" rx="16" fill="#334155" stroke="#38BDF8" stroke-width="2"/>
  <text x="960" y="760" text-anchor="middle" fill="#38BDF8"
        font-size="40" font-family="sans-serif" font-weight="bold">Database</text>
</svg>

Workflow

In the Step 1.5 slide spec JSON, specify the output path in the image field
- Nano Banana: specify a .png path
- SVG: specify a .svg path (auto-converted)
Generate images in this step
Images are automatically inserted when slides are generated in Step 2

Step 2: Slide Creation

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh create_slides.py output/slide_spec.json output/presentation.pptx

To use a custom template:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh create_slides.py output/slide_spec.json output/presentation.pptx --template path/to/template.pptx

After generation, ask the user to review the PPTX.

Step 3: Talk Script Update (Optional)

If notes are already included from Step 1.5, this step can be skipped. Execute when the user wants to change the writing style or regenerate scripts.

What you (Claude) should do

Read the current notes:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh write_talk_script.py output/presentation.pptx --read

Refer to references/talk-styles.md and regenerate scripts for all slides in the user's preferred style
- Styles: casual (default) / keynote / formal / humorous
Save as a JSON array and write to the PPTX:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh write_talk_script.py output/presentation.pptx --notes-json output/notes.json --output output/presentation.pptx

Step 4: Speech Synthesis

Confirm with the user

TTS engine: edge (default) / voicevox
Voice: default is ja-JP-NanamiNeural (edge), 1 = Zundamon (voicevox)

Step 4a: TTS Script Export

Export speaker notes as a text file for text-to-speech. When a pronunciation dictionary (--dict) is specified, common mispronunciation patterns are automatically replaced.

With pronunciation dictionary (recommended):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh export_tts_script.py output/presentation.pptx output/tts_script.txt --dict ${CLAUDE_SKILL_DIR}/assets/pronunciation_dict.tsv

Without pronunciation dictionary:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh export_tts_script.py output/presentation.pptx output/tts_script.txt

Step 4b: Review and Edit Pronunciation Text

Present the exported script file contents to the user. Prompt them to check for items not auto-corrected by the dictionary, or project-specific terminology.

Examples auto-corrected by dictionary:

"生成" → "せいせい", "Claude" → "クロード", "LLM" → "エルエルエム"

Cases requiring manual correction:

Project-specific abbreviations and product names
Technical terms not registered in the dictionary

If the user decides no corrections are needed, proceed to Step 4c (synthesize without --script, as before). If entries should be added to the dictionary, update pronunciation_dict.tsv.

Step 4c: Script Execution

With a corrected script file (--script):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine edge --script output/tts_script.txt

Without corrections (synthesize directly from PPTX notes, as before):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine edge

When using VOICEVOX (VOICEVOX Engine must be running beforehand):

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh synthesize_audio.py output/presentation.pptx output/audio/ --engine voicevox --script output/tts_script.txt

TTS Failure Handling

If the TTS API fails for some or all slides, failed slides are skipped and audio files are generated only for successful ones (partial success).

The script reports failure count to stderr and exits normally
Slides without audio will auto-advance with default timing (3 seconds) in Step 6
After TTS recovery, re-running Step 4c will generate audio for all slides
The PPTX from Step 2 remains usable as-is (presentation is possible without audio)

Step 5: Audio Embedding

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh embed_audio.py output/presentation.pptx output/audio/ output/presentation_with_audio.pptx

Step 6: Slideshow Auto-Play Configuration

Slides with audio: auto-advance after audio playback + buffer (default 1 second)
Slides without audio (title, section headers, etc.): auto-advance with fixed time (default 3 seconds)

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_slideshow.py output/presentation_with_audio.pptx output/presentation_final.pptx

To adjust display timing:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_slideshow.py output/presentation_with_audio.pptx output/presentation_final.pptx --silent-duration 5000 --audio-buffer 2000

User Guidance (Required)

After completing the slideshow configuration, always communicate the following notes to the user:

Pre-Slideshow Checklist

For PowerPoint (Windows / macOS): Verify that "Use Timings" is checked in the slideshow settings.

Windows: "Slide Show" tab → "Set Up Slide Show" → Check "Use Timings"

macOS: "Slide Show" menu → "Set Up Show..." → "Options" → Check "Use Timings"

If this is unchecked, auto-advance and audio auto-play will not work.

For LibreOffice Impress: Auto-advance does not work. Please advance slides manually during the slideshow (click or arrow keys). Audio will auto-play on each slide.

Step 7 Guidance (Required)

Following the Step 6 guidance, always confirm the following:

This presentation can also be exported as an MP4 video. This is useful for posting to YouTube or sharing in environments where PowerPoint is not available. However, additional tools (LibreOffice, ffmpeg) need to be installed. If you don't need video, the process is complete here. Would you like to create a video as well?

If the user responds with "Not needed", "No thanks", etc., skip Step 7 and end the pipeline.

Step 7: MP4 Video Export (Optional)

Generate an MP4 video from the PPTX and audio files. Execute when you want to distribute or share the slideshow as a video.

Confirm with the user

When the user responds "Create the video" to the Step 7 guidance, confirm the following:

Video frame rate (default: 30fps)
Display duration for slides without audio (default: 3 seconds)

Prerequisites

Video generation requires additional tools. Availability is shown during setup.sh execution.

Tool	Purpose	Required
LibreOffice	PPTX → PDF → PNG rendering	Yes
ffmpeg	Video clip generation & concatenation	Yes
pdftoppm (poppler-utils)	PDF → PNG conversion (high quality)	Recommended (ffmpeg can substitute)

If not installed:

# Ubuntu/Debian
sudo apt install libreoffice ffmpeg poppler-utils

# macOS
brew install libreoffice ffmpeg poppler

Note on snap version of LibreOffice: Use an output directory under $HOME. Writing to /tmp etc. may fail due to sandbox restrictions.

Script Execution

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_video.py output/presentation_final.pptx output/audio output/presentation.mp4

With options:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh make_video.py output/presentation_final.pptx output/audio output/presentation.mp4 --fps 30 --silent-duration 3.0

Automatic Validation

After generation, the following checks are automatically performed:

Check Item	Details
File existence	MP4 file exists and size > 0
ffprobe read	File is not corrupted
Codecs	Video: H.264 / Audio: AAC
Resolution	Width and height are even numbers (libx264 compatible)
Duration	Positive value

If validation fails, error details are displayed. Use --skip-validation to disable.

User Guidance

After video generation is complete, communicate the following:

About the Video Output

presentation.mp4 has been generated

Display time for each slide is based on the actual audio file duration

Slides without audio display for 3 seconds by default

Playable with common video players (VLC, QuickTime, etc.)

Format Conversion (Optional)

If ODP format is needed:

bash ${CLAUDE_SKILL_DIR}/scripts/run.sh convert_format.py output/presentation.pptx --outdir output/

Prerequisite: LibreOffice must be installed.

Cross-Platform Compatibility Notes

macOS PowerPoint Compatibility

PPTX files generated by Relief Pitcher AI work on both macOS and Windows PowerPoint as well as LibreOffice Impress. Note the following technical constraints.

Audio Auto-Play (Step 6)

Method	macOS	Windows	LibreOffice
`p:cmd type="call" cmd="playFrom(0)"`	Yes	Yes	Yes
`p:audio > p:cMediaNode`	No (treated as corrupt)	Yes	Yes

macOS PowerPoint treats p:audio > p:cMediaNode (OOXML media node method) as a corrupted file
Instead, use p:cmd type="call" cmd="playFrom(0)" (command animation method)
Specify nodeType="afterEffect" + grpId="0" for "auto-execute after previous animation"
This method also works correctly on Windows PowerPoint / LibreOffice Impress

Audio Icon (Step 5)

Icon	macOS	Windows
32x32 visible icon	Yes, displayed	Yes, displayed
1x1 transparent PNG	No, invisible	Yes, auto-replaced with speaker icon

macOS PowerPoint displays transparent icons as-is (Windows auto-replaces with a speaker icon)
Therefore, a 32x32 visible speaker icon is used when embedding audio
The icon is dynamically positioned at the bottom-right corner based on slide dimensions (0.25" margin from right and bottom edges)
Correctly positioned at bottom-right for both 4:3 and 16:9 aspect ratios

Auto-Advance

Method	macOS	Windows	LibreOffice
`p:transition advTm`	Yes	Yes	No (not supported)

LibreOffice Impress does not support auto-advance. Even with timing settings (advTm, mainSeq.dur) correctly configured in the PPTX, LibreOffice ignores them — this is a known limitation (Bug 101527)
To play in LibreOffice, advance slides manually or manually set "auto-transition" time for each slide in LibreOffice's "Slide Transition" panel
Auto-advance works correctly in PowerPoint (Windows / macOS)

Reference

references/pptx-guide.md: Details on layouts and placeholders
references/talk-styles.md: Talk script style definitions
references/tts-plugins.md: How to add TTS plugins

relief-pitcher-ai

同仓库更多 Skills

同仓库更多 Skills

Relief Pitcher AI — Auto-Generate Presentation Skill

Overview

Help Display

Prerequisites

Common Pattern for Script Execution

Workflow Selection (Step Resume Support)

Video Output Confirmation

Step Resume / Interruption

Step 1: Outline Generation

What you (Claude) should do

Script Execution

Step 1.5: Content Enrichment

What you (Claude) should do

Diagram Density Levels

JSON Format

Script Execution

Step 1.7: Image Generation (Optional)

Choosing a Method

SVG-suited (Claude generates directly)

Nano Banana-suited (generated via Gemini API)

Confirm with the user

Nano Banana (Gemini Image Generation API)

SVG Generation (No API Required)

Procedure

SVG Sizes

SVG Font Size Requirements

Color Scheme by Template

SVG Pattern Examples

Workflow

Step 2: Slide Creation

Script Execution

Step 3: Talk Script Update (Optional)

What you (Claude) should do

Step 4: Speech Synthesis

Confirm with the user

Step 4a: TTS Script Export

Step 4b: Review and Edit Pronunciation Text

Step 4c: Script Execution

TTS Failure Handling

Step 5: Audio Embedding

Script Execution

Step 6: Slideshow Auto-Play Configuration

Script Execution

User Guidance (Required)

Step 7 Guidance (Required)

Step 7: MP4 Video Export (Optional)

Confirm with the user

Prerequisites

Script Execution

Automatic Validation

User Guidance

Format Conversion (Optional)

Cross-Platform Compatibility Notes

macOS PowerPoint Compatibility

Audio Auto-Play (Step 6)

Audio Icon (Step 5)

Auto-Advance

Reference

Relief Pitcher AI — Auto-Generate Presentation Skill

Overview

Help Display

Prerequisites

Common Pattern for Script Execution

Workflow Selection (Step Resume Support)

Video Output Confirmation

Step Resume / Interruption

Step 1: Outline Generation

What you (Claude) should do

Script Execution

Step 1.5: Content Enrichment

What you (Claude) should do

Diagram Density Levels

JSON Format

Script Execution

Step 1.7: Image Generation (Optional)

Choosing a Method

SVG-suited (Claude generates directly)