Tech Tues Ai Chat Export Tool

Last week I shared how I reverse-engineered a platform to extract 20,000+ AI chat messages.

This week, I'll talk about what I built to actually analyze them.

The problem: Raw data isn't insight. I had 20k+ messages in a JSONL file. I needed tools to study patterns, compare prompts to responses, and understand what was actually working in my creative workflow.

The platform wasn't going to build that for me, so I built it myself.

The viewer:

FastAPI backend, React frontend, all running locally.

The chat reader handles 20,000+ messages with virtualized scrolling. Messages are speaker-differentiated, action text parses into styled formatting, and time gaps show as dividers so I can see session boundaries.

The dashboard:

Daily message density, activity heatmaps showing when I tend to work, response time distributions, session detection. I can filter by date range and see how my workflow changes over time.

The research tools:

This is where it gets useful.

A prompt-response pair browser that lets me study what I sent versus what I got back. A prompt patterns library where I catalog structural approaches and measure which ones produce better output. A reroll browser that shows every time I regenerated a response, with all candidates side by side.

I also built an annotation system with tags, flags, and freeform notes stored in a SQLite sidecar so the source data stays untouched. Over 2,200 messages annotated so far.

The takeaway:

If you use AI for creative work, your conversation history is a prompt engineering dataset hiding in plain sight.

Platforms won't build research tools for you. But with a weekend of building, you can turn chat logs into a lab.

---

Adjust as you build this week. Spots you might want to swap in fresh details:

• Specific numbers (messages annotated, patterns identified)
• A concrete finding from the research tools
• Any new features you add