In Simple Terms
How do you build a system with 76,000 lines of code and 18 services? Not alone, and not in the traditional way. This paper documents how a human developer and AI coding assistants (Claude Code) co-developed Frank's architecture through iterative cycles of building, testing, discovering bugs, and redesigning.
Development Pattern
- Human designs the concept. "Frank needs a sensory gating system like the thalamus."
- AI implements the first draft. Claude Code writes the initial 570 lines.
- Testing reveals unexpected behavior. Frank fixates on "whispers of the Architecture Bay" for 5 hours.
- Joint diagnosis. Human and AI analyze the logs together, identify the root cause.
- AI implements the fix. 7 targeted patches to the anti-rumination system.
- Repeat.
Key Discoveries Made Through Codevelopment
- The regex co-author bug (v6 of IAPT) — discovered through adversarial testing that no human reviewer would have caught
- The mood decay non-cumulative bug — E-PQ was re-reading the base value every cycle instead of tracking deltas
- The watchdog cascade — services trying to restart non-existent services, creating a chain reaction
- The negative instruction priming — training data with "NEVER be nihilistic" actually made the model more nihilistic
Why It Matters
The paper provides evidence that human-AI codevelopment isn't just "AI writes code, human reviews." It's a genuine collaborative process where each party contributes capabilities the other lacks. The human provides architectural vision and judgment. The AI provides implementation speed and exhaustive testing.