1.9 KiB
The Verification Loop
Boris Cherny (the creator of Claude Code) says verification loops deliver 2-3x quality improvement. That's not a small number.
The hard truth from Reddit: "Claude can write code all day but can't click a button to see if it works." AI-generated code is NOT optimized for security, performance, or scalability by default. You have to explicitly prompt for these and always verify.
The PM Angle
You are the QA layer. Every output needs human eyes before it ships. This isn't new for PMs — we already review designs, specs, and code. The difference is the volume and speed of output.
Debugging Patterns That Work
Blind paste — copy error messages directly into Claude without reading them yourself. Claude is remarkably good at diagnosis from raw errors. 0xDesigner: "I got another error and just pasted the error without reading." Claude diagnosed a missing gas token from a raw wallet error.
Screenshot debugging — paste screenshots of UI bugs directly. Claude's vision is powerful for visual debugging. "A single screenshot can replace three paragraphs of explanation."
The re-prompt — if output seems wrong, use Esc-Esc to go back and rephrase slightly. Often a slight rephrasing yields vastly superior results.
The TDD Loop for AI
Powerful pattern from a practitioner with 6 months of daily AI pairing:
- AI writes a failing test
- Human reviews the test
- AI writes the implementation
- Test passes → move on
This enforces requirements and ensures correctness. Even without formal TDD: always ask Claude to verify its own work.
Connection to Other Notes
- The Junior Employee Mental Model — you wouldn't trust a junior's work without review
- Plan Mode — verification starts before execution
- The Death Spiral — skipping verification is how you get there
- Boris Cherny's "Grill me on these changes" prompt — make Claude critique its own work. See Custom Commands for how to automate this.