Slide Deck Generation
Generate professional presentation slides using AI through a 4-step workflow:
- Initialize Task - Create task workspace with unique ID
- Generate Outline - Create structured outline (YAML)
- Generate Images - Run script to create slide images
- Assemble PDF - Combine images into final PDF
Prerequisites
API Key - Set in .env file at skill root (auto-loaded by script):
GEMINI_API_KEY=your-api-keyDependencies:
pip install -r requirements.txtWorkflow
Step 0: Initialize Task Workspace
When the user starts creating a PPT, immediately generate a Task ID and create a workspace under the project root directory:
$TASK_ID = (Get-Date -Format "yyyyMMdd_HHmmss") + "_" + -join ((1..6) | ForEach-Object { '{0:x}' -f (Get-Random -Maximum 16) })
New-Item -ItemType Directory -Force -Path "{workspace}/output/$TASK_ID"{workspace} refers to the user's project root directory, not the skill directory.All files will be saved under {workspace}/output/<task-id>/.
Step 1: Generate Outline
You are a world-class presentation designer and storyteller.
First, ask the user to choose:
| Option | Values | Description |
|---|---|---|
| Format | detailed / medium / presentation | Detailed (standalone reading) / Medium (speech + reading) / Minimal (speech-only aid) |
| Length | short / default | Short (5-10 slides) / Default (10-20 slides) |
| Language | auto | Auto-detect from user's prompt language; user may also specify explicitly |
| Other | Target audience, style preferences |
Then generate outline as YAML, save to {workspace}/output/<task-id>/outline.yaml:
style_instructions:
format: detailed # detailed | medium | presentation
length: default # short | default
language: <auto-detected or user-specified language>
design_aesthetic: <design aesthetic description>
background_color: "#F8F7F5"
primary_font: <heading font>
secondary_font: <body font>
color_palette:
primary_text: "#2F3542"
primary_accent: "#007AFF"
visual_elements: <visual element style description>
slides:
- slide_number: 1
slide_type: cover # cover | content | closing
narrative_goal: <narrative goal>
key_content:
title: <title>
subtitle: <subtitle>
body: |
<body text, adjust density based on format>
visual:
description: <detailed visual description for the image generation model>
elements: [<element1>, <element2>]
layout:
composition: <composition description>
hierarchy: <visual hierarchy>
_metadata:
task_id: "<task-id>"
topic: <topic>
format: detailed
slide_count: 10Critical Rules:
- Slide 1 must be a cover page - Poster-style layout, bold typography, full-bleed imagery
- Last slide must be a closing page - Meaningful visual summary, never end with "Any questions?" or "Thank you"
- Never exceed 20 slides
- Avoid "Title: Subtitle" format - Use narrative thematic sentences to connect slides
- Avoid AI cliché patterns - Never use phrases like "It's not just X, it's Y"
- Visual descriptions must be thorough - Detailed enough for an image generation model to understand
- Adjust content density based on format
Show the outline to the user for approval before proceeding.
For format-specific outline examples and style templates, see references/outline_generation.md.
Step 2: Generate Slide Images
After user approves the outline:
python scripts/generate_slides.py {workspace}/output/<task-id>/outline.yaml \
--output-dir {workspace}/output/<task-id>/slidesCommon options:
| Argument | Default | Description |
|---|---|---|
--resolution | 4K | 1K, 2K, 4K |
--max-concurrent | 4 | Concurrency (reduce to 2 if rate-limited) |
--proxy | None | HTTP proxy URL |
--slides | all | Specify slides: 1,3,5 or 2-5 |
Step 3: Assemble PDF
python scripts/assemble_pdf.py {workspace}/output/<task-id>/slides/ {workspace}/output/<task-id>/presentation.pdfOptional: --compress, --watermark logo.png, --watermark-position bottom-right, --watermark-opacity 0.5
Task Workspace Structure
{workspace}/
└── output/
└── 20240111_143052_a1b2c3/
├── outline.yaml
├── slides/
│ ├── slide_01.png
│ └── ...
└── presentation.pdfError Handling
- Rate Limits: Reduce
--max-concurrentto 2 - Failed Slides: Use
--slidesto regenerate specific ones - Connection Error: Check proxy settings, script auto-retries (max 3, exponential backoff)
For detailed script API reference, see references/REFERENCE.md.