Developer switches from Claude Code to Pi over prompt overhead

The move to Pi

A developer who blogs at caleb.software recently shared their decision to retire Claude Code as their primary coding harness. They chose to switch to Pi, which is an open-source alternative licensed under the MIT license. The move was described as a result of an accumulation of small frustrations with the official Anthropic tool.

The author expressed that while Anthropic has built a cool piece of software, the current experience lacks a clear vision or direction. They noted that Pi provides a benefit of hindsight, allowing users to see what works and what does not in a more streamlined package. The developer's post gained significant attention on Hacker News, reaching the front page in November.

Pi is described by its website as a minimal agent harness that provides powerful defaults while skipping complex features like sub-agents. It allows for high customization through themes, prompt templates, and skills. Because it supports over 15 providers, users can still utilize Claude models within the Pi framework.

Prompt overhead and costs

A major point of contention for the developer was the frequency of updates to Claude Code. They counted 244 releases for the tool this year, noting that more than three-quarters of those updates changed the system prompt or tool descriptions. These changes can substantially alter how the underlying large language model performs during a session.

The author highlighted specific technical overhead, noting that the default Bash tool's description in Claude Code is almost 3,000 tokens. They argued that Claude Code maintains significantly more overhead than Pi without providing a clear benefit to the user experience. This technical weight can impact the efficiency of the interaction between the user and the AI.

In a footnote, the developer estimated that trimming 1,000 tokens from these descriptions could save millions of dollars in inference costs annually. This calculation was based on an assumption of one million daily users making 100 million requests a day. While these are the author's own estimates, they underscore a growing concern regarding the hidden costs of large system prompts.

Integration and workflow

Beyond the underlying prompts, the developer cited several practical issues with the Claude Code VS Code extension. They noted that only one session can be connected to the editor at a time, which limits multitasking. Additionally, they felt the extension lacked a custom status line and tended to mix subagent output into the main thread.

The developer also struggled with maintaining different file formats, specifically having to manage CLAUDE.md files alongside AGENTS.md files. This created a manual overhead that made the workflow less seamless than they desired for a professional coding environment. These friction points contributed to the decision to seek a more streamlined harness.

Interestingly, Anthropic announced support for AGENTS.md files in Claude Code just one day after the developer decided to leave. While the official documentation now states that Claude can read these files in place of CLAUDE.md, the developer's experience shows how even small gaps in official tools can drive users toward open-source alternatives.

The harness vs the model

This story highlights a distinction between the AI model itself and the harness that wraps it. The harness is the software layer that compresses a team's feedback, the model's raw capabilities, and specific instructions into a replicable tool. The developer's switch was not a rejection of Claude's models, but rather a choice between two different ways to package those models.

By using Pi, the developer can still access Anthropic's models while benefiting from a minimal system prompt and a different set of features. Pi's approach of skipping sub-agents and plan mode in favor of extensions allows users to build a more tailored experience. This reflects a broader trend where the