Boost coding agent quality with fan-out criticism and self-rating rules
jbarbier · x · 2026-08-16
The author shares two new rules added to an open-source CLAUDE.md file to improve coding agent performance:
- Fan-out + Harsh Critic: Inspired by game-replication loops, this generates 2-3 competing variants for every feature. A separate critic judges them blind, rejects by default, and loops until one passes. Bug fixes include an "attacker" trying to re-break them.
- Self-rating: After every task, the agent scores its work (1-10) and answers honestly: "Am I proud of this?" If no, it must name the gap, fix it, and loop. Inflating scores to exit is banned.
More from coding & agent
- Hello world in 4 minutes: Generative UI dramatically speeds up development — rseroter · 2026-08-16
- GitHub 19.3k Stars: A Style Guide That Lets Claude Draw Designer-Grade Diagrams — socialwithaayan · 2026-08-16
- Monitor AI Agents via iPhone Home Screen Widgets — Dense-Map-406 · 2026-08-16
- Building a Complete Evaluation and Experimentation Pipeline for Agents — blaizedsouza · 2026-08-16
- The Full Agent Lifecycle: From planning to continuous monitoring for production AI — blaizedsouza · 2026-08-16
- Google releases free 2-hour course on graph engineering for multi-agent systems — blaizedsouza · 2026-08-16