A single pass with an AI coding assistant misses things. That's not a flaw in the model - it's how probability-based text generation works. Every time you run the same prompt, you get slightly different results. Stephan Schmidt, writing on his AmazingCTO blog, argues that Claude Code's /loop command turns this weakness into a strength.
The /loop command lets you run any prompt on a recurring interval - say, every 10 minutes. Schmidt tested it with a bug-hunting prompt: /loop 10min "deep search src/ for bugs. add to potential_bugs.md". The first pass found 12 bugs. The second found 5 more that the first missed. By run five, no new bugs appeared - the analysis had converged on a thorough result.
The principle behind this comes from a 2022 Google Research paper on "self-consistency," which showed that running a prompt multiple times and aggregating the results produces more reliable outputs than any single run. Each pass explores a slightly different path through the model's probability space, catching things the previous passes skipped.
This isn't limited to bug hunting. The same technique applies to security audits, architecture reviews, or any task where completeness matters more than speed. Instead of trusting one AI pass to catch everything, you let the model take multiple shots and compile the results.
The /loop feature has been in Claude Code for a while, but using it specifically for convergence-based analysis is a neat trick. If you're already paying for Claude Code, it costs nothing extra to run - just the token usage from repeated passes. For codebases where missed bugs have real consequences, five passes at 10-minute intervals is a small time investment for meaningfully better coverage.