Goodbye Claude, Hello Pi

A longtime user of Claude Code has decided to switch to Pi, citing a series of frustrations and issues with the former. Claude Code, introduced over a year ago in the Claude Pro subscription plan, was initially met with enthusiasm for its ability to generate code from a single prompt. However, the author found many small annoyances that eventually led to its retirement as their primary coding tool.

One major frustration was the AGENTS.md feature, which the author believes was a misguided attempt to improve user retention. While having scattered CLAUDE.md files in GitHub repositories was great for marketing, the author found it inconvenient to manage the different formats. They also struggled with the limitation of only being able to connect one session to the IDE at a time, making it difficult to work on multiple tasks simultaneously.

Another issue was the frequent changes to the system prompt and tool descriptions, which often altered the LLM's behavior and contradicted user prompts or repository conventions. Despite these changes, the default Bash tool's description remained excessively long and included irrelevant information, while the model's performance remained unchanged.

The author also expressed concerns about the rushed development of Claude Code, with over 244 releases in a single year. Many of these updates focused on adding new features rather than addressing existing issues or maintaining stability. This, combined with the lack of clear direction and a strong vision for the product, left the author feeling that Claude Code was a project without clear priorities.

In contrast, the author found Pi to be a more stable and customizable option. The transition required some initial effort, but once settled, the author has been pleased with the performance and ease of use. They have even started creating extensions to support their specific needs, hoping to share them in the future.

For those facing similar frustrations with Claude Code, the author encourages exploring open-source alternatives. They highlight the potential cost savings from trimming unnecessary tokens from the model's descriptions and suggest that such improvements should be prioritized by engineers. Overall, while Anthropic has made significant strides in LLM coding harnesses, the author believes that a stronger focus on quality, stability, and user needs is essential for long-term success.