Generating slides

A multimodal agent recipe. Given a topic, the model uses the Model API's built-in web search tool (tools=[{"type": "web_search"}]) to research it via the Responses API, synthesizes the findings into a structured outline, and generates a reveal.js slide deck. The model controls both content and design.

How it works

topic -> Model API web search (3-5 queries) -> research brief -> JSON outline -> model-designed HTML

1. Research: The model uses the Model API's web_search tool to search the web from multiple angles. The model autonomously decides what queries to run.
2. Outline: The model organizes the research into a structured JSON outline (slide titles, bullet points, code blocks, citations). The outline is normalized to handle key name variance across model outputs.
3. Design: The model generates the full HTML presentation - structure, CSS, and content - with creative control over the visual design.

Setup

bash
# With uv
uv venv && uv pip install -r requirements.txt
​
# Without uv
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Quick start

bash
export MODEL_API_KEY="your-key"
​
# Generate slides on any topic
python research_to_slides.py "AI Agents in Production"
​
# Open the result
open ai_agents_in_production_slides.html

The pipeline produces two files:

• *_slides.html - the presentation (open in any browser, arrow keys to navigate)
• *_slides_stats.json - pipeline metrics: search queries, token usage, timing, sources

See examples/ for a pre-generated example deck, stats, and slide screenshot.

Files

Environment variables

• MODEL_API_KEY - required for all pipeline runs
• META_MODEL - model name, defaults to muse-spark-1.3
• META_BASE_URL - API endpoint, defaults to https://api.meta.ai/v1