Audio-Multimodal-Formatting-Prompts

Generate creative system prompts for audio-to-text transformation workflows using AI.

Overview

This project generates system prompts that instruct audio multimodal models to transform voice recordings into formatted text in creative and diverse ways.

Quick Start

Setup

# Install dependencies
pip install -r requirements.txt

# Or use uv
uv pip install -r requirements.txt

Running Generations

Run All (Recommended)

./run_all.sh

Runs all three strategies sequentially with appropriate delays between runs.

Individual Runs:

Run 1: Batch Generation (10 prompts)

source .venv/bin/activate
python generate_prompts.py 1

Uses system-prompts/1.md to generate 10 prompts in a single API call.

Run 2: Individual Generation (50 prompts)

source .venv/bin/activate
python generate_prompts.py 2

Uses system-prompts/2.md to generate 50 unique prompts (50 individual API calls with 1s delay).

Run 3: Large Batch Generation (50 prompts)

source .venv/bin/activate
python generate_prompts.py 3

Uses system-prompts/3.md to generate 50 diverse prompts in a single API call with maximum creative diversity.

Options

# Use a different model
python generate_prompts.py 1 --model anthropic/claude-3.5-sonnet

# Adjust delay between requests (for run 2)
python generate_prompts.py 2 --delay 2.0

Available Models

• google/gemini-2.5-flash (default - latest, fast, good reasoning)
• google/gemini-2.5-flash-lite (even faster, lighter)
• google/gemini-2.0-flash-exp:free (free tier)
• anthropic/claude-3.5-sonnet (higher quality, costs credits)

Output Structure

Outputs are organized into two directories:

runs/ - Programmatic Data

Contains JSON files with structured data for programmatic access:

• run1_batch_10_{timestamp}.json
• run2_individual_50_{timestamp}.json
• run3_batch_50_{timestamp}.json

outputs/ - Human-Readable Data

Organized by run with markdown and PDF formats:

outputs/
├── run1/
│   ├── markdown/
│   │   ├── {run_name}_{timestamp}_all.md  (consolidated)
│   │   └── prompts/                         (individual files)
│   │       ├── 001_prompt-name.md
│   │       ├── 002_prompt-name.md
│   │       └── ...
│   └── pdf/
│       └── {run_name}_{timestamp}.pdf
├── run2/
│   └── ...
└── run3/
    └── ...

Each Prompt Contains:

• name: Short descriptive name (e.g., "Meeting Minutes - Action Items")
• description: What the transformation does and when to use it
• prompt: Complete system prompt ready for audio multimodal models

Generating PDFs

After running the generators, create PDFs with:

source .venv/bin/activate
python generate_pdfs.py

This creates a formatted PDF for each run with:

• Cover page with run metadata
• One prompt per page with title, description, and full prompt
• Professional formatting for easy reading and printing

Consolidating All Runs

To create a single consolidated view of all prompts across all three runs:

source .venv/bin/activate
python consolidate_runs.py

This creates:

• outputs/consolidated/markdown/all_prompts_{timestamp}.md - All 118 prompts organized by run
• outputs/consolidated/pdf/all_prompts_{timestamp}.pdf - Single PDF with all prompts and run indicators in footers

The consolidated output includes:

• Run descriptions explaining each generation strategy
• All prompts organized by run number
• Footer on each prompt page indicating which run it came from

System Prompts

• system-prompts/1.md: Batch generator (10 prompts)
• system-prompts/2.md: Single-turn generator (1 prompt per call)
• system-prompts/3.md: Deep ideation generator (50 prompts with maximum diversity)

Configuration

Add your OpenRouter API key to .env:

OPENROUTER_API_KEY=your_key_here