Episode 41 - Reverse Engineering Legacy Code With AI Browser Automation

Episode Summary

Martin shares his experience running Claude Code on a 1.2-million-line codebase for 40 hours (~$500 in API costs), then discovering that browser automation via MCP produced far better event models. The full strategy: overlay event models from UI automation, database deltas, code analysis, and documentation.

Main Discussion Points

• 40-Hour AI Legacy Analysis — Martin ran Claude Code on a 1.2M-line codebase for ~40 hours, burning ~40 million tokens (~$500); the code-based event models were okay but not great
• Browser Automation as the Better Approach — Pointing Claude Code via MCP at the legacy UI and saying “explore this and build an event model” dramatically outperformed static code analysis
• UI Screenshots → Event Model — Feed browser automation screenshots directly into the event model: commands from buttons, read models from data displays, events inferred from state changes
• Database Delta Watching — Sandwich UI automation with database snapshot diffing: if a “navigation” click causes DB writes, it reveals a hidden command disguised as a read
• Multi-Angle Legacy Analysis — The most robust strategy overlays event models from: UI automation, database deltas, codebase analysis, and documentation; gaps reveal technical debt
• Documentation Correlation — Using Git history and wiki page timestamps to determine whether documentation predates or follows implementation; stale docs actively mislead
• Local GPU Inference Economics — Adam runs smaller models on his own GX10 hardware; Martin has three Spark machines; local vs. cloud inference tradeoffs
• Event Model JSON as the Specification — Reiterating that the Miro JSON export is already the complete AI-ready specification; no translation layer needed
• Work Shifting to Higher Abstraction — Industry catching up to what event modeling practitioners knew: “the spec is the thing, not the code”
• Open Agent Orchestration — Discussion of open-source agent orchestration frameworks; both using MCP browser automation

Browser Automation as the Better Approach

“I just pointed Claude Code using an MCP server to this UI. I didn’t tell it anything. Just said: go to this UI, find out what this thing is doing, and build an event model from that. It worked like a charm.” - Martin

The Spec Is the Thing

“Code is no longer the thing. It’s the spec. Oh, really? Interesting. Thanks for telling us what we knew for many many years.” - Adam

Lessons No Longer Apply

“The lessons you learned from before AI no longer apply — just like lessons you learned for CRUD don’t apply to event sourcing. It’s the same thing.” - Martin

Key Takeaways

1. Browser automation beats static code analysis for legacy — Pointing an AI agent at the UI dramatically outperforms 40 hours of static code analysis
2. Database deltas reveal hidden commands — Snapshot diffing exposes navigation clicks that secretly write to the database
3. Multi-angle overlay produces the most complete event model — UI automation + database deltas + code analysis + documentation fill each other’s gaps
4. Stale documentation actively misleads — Git history and timestamps reveal whether docs predate or follow implementation
5. ~$500 for 1.2M LOC analysis is affordable — The economics of AI-powered legacy analysis are increasingly viable
6. Event model JSON is already AI-ready — The Miro JSON export is the complete specification; no translation layer needed
7. Local GPU inference is becoming viable — GX10 hardware runs smaller models competitively against cloud inference
8. The spec is the new source of truth — Industry is catching up to what event modeling practitioners knew for years