Ponytail vs. a good AGENTS.md: What we found on our real codebase (docspring.com)

A recent evaluation of the Ponytail plugin for AI coding agents, particularly Claude Code and Opus, revealed that while it significantly reduces code volume and development costs—writing 53% fewer lines and 26% cheaper—it may not be the best fit for all codebases, particularly complex ones. The evaluation was based on a real codebase and involved analyzing over 10,000 messages, categorizing complaints about code, and running multiple benchmarks. Despite Ponytail's efficiency in producing leaner code, it sometimes fell short on completeness and could increase the complexity of tracking revisions, particularly in environments with established operational rules. In response to the findings, the team chose not to implement Ponytail in its entirety but instead adopted key principles from it to enhance their existing AGENTS.md framework. They emphasized creating the smallest complete changes possible, introducing a more robust review process for identifying unnecessary code, and ensuring thoroughness in change audits. These adaptations are designed to maintain their code's quality while addressing the overproduction of unnecessary features, thereby streamlining their development process without compromising on necessary adjustments. The team is sharing their newfound strategies and principles on GitHub, providing a resource for others looking to refine their AI-assisted coding practices.