Today I learned how to convert Firefly meeting transcripts (which export as HTML) into clean, simplified markdown that’s much easier for LLMs to process.

The Problem

Firefly transcripts come as HTML files packed with inline data:image tags (profile pictures, avatars), emoji characters, and HTML formatting. This bloat makes them slow to process and wastes tokens when feeding them to LLMs.

The Solution: Two Scripts

convert_html.py and remove-images.py

Step 1: HTML to Markdown with Pandoc

Use the convert_html.py script which wraps Pandoc for clean conversion:

# Basic conversion
python scripts/convert_html.py transcript.html

# Strip all HTML tags for cleaner output
python scripts/convert_html.py transcript.html --clean

Prerequisites: Install Pandoc first with brew install pandoc.

The --clean flag uses Pandoc’s -t gfm-raw_html option, which outputs GitHub Flavored Markdown while stripping any residual HTML tags. This produces much cleaner output for LLM consumption.

Step 2: Remove Embedded Images and Emojis

The converted markdown may still contain embedded data:image URIs and emoji characters. Clean these with remove-images.py:

# Remove data:image tags and emojis (default)
python scripts/remove-images.py transcript.md

# Preview what would be removed
python scripts/remove-images.py transcript.md --dry-run -v

# Keep emojis if needed
python scripts/remove-images.py transcript.md --keep-emojis

Key options:

• --all-images: Remove all images, not just data URIs
• --no-backup: Skip creating a .bak file
• -d, --dry-run: Preview changes without modifying the file

Why This Matters

Firefly transcripts can contain hundreds of embedded profile images as base64 strings. A single avatar might be 50KB+ of text. Removing these drops the file size dramatically—often by 90% or more—making LLM processing faster and cheaper.

The cleanup also removes Unicode emoji characters that can cause encoding issues with some tools.

Complete Workflow

# Convert HTML to Markdown
python scripts/convert_html.py meeting.html --clean

# Clean the result
python scripts/remove-images.py meeting.md --no-backup

# Now feed to your LLM!