Developer Tools / ARTICLE

Goodbye Claude, Hello Pi

A longtime Claude Code user explains his switch to Pi, citing IDE integration limits, frequent releases that changed system prompts, and bloat in the default Bash tool description.

Developer Caleb describes leaving Claude Code after more than a year of use, moving his main coding harness to Pi. He writes that Anthropic announced support for AGENTS.md files one day after his decision, an attempt to keep him, but he calls it ‘too little, too late.’

IDE Integration Gaps

Caleb used Claude Code inside VSCode, where the harness can connect to the IDE. Highlighting text in a file automatically includes it in the context for questions or change requests during code review. However, only one session can be connected to the IDE at a time. If he opened a second tab to start another task while the first was running, that second session would not receive text selection events from the IDE. He says the issue was reported as far back as December 2025, but it was closed as a duplicate of an unrelated bug.

The official Claude Code VSCode GUI extension allows multiple tabs and text selection, but it has its own problems. Half of the built-in slash commands do not work in the GUI. There is no support for a custom status line. Subagent output messages are layered into the main chat thread in chronological order, which becomes messy when more than one subagent is running.

Caleb notes that after installing Pi, he wrote an extension to handle text selection across multiple open tabs in about 20 minutes.

Release Velocity and Token Overhead

He points out that Anthropic has packed more features into Claude Code without much care for quality or stability. Dynamic workflows, agent teams, and ultrareviews, he writes, are great for helping enterprise customers burn more tokens each day, but not things he wants or needs.

These releases also disrupt his personal setup. He counts 244 releases of Claude Code this year, with over three quarters changing the system prompt or tool descriptions. Such changes can substantially alter how the LLM performs, especially when new instructions contradict user prompts or repository conventions.

Despite trimming 80 percent of the system prompt, Claude Code still carries far more overhead than Pi, he says. The default Bash tool description is almost 3000 tokens, most of it about using git and opening pull requests on GitHub. He argues the models already know how to do this without detailed instructions. He includes a rough estimate: assuming 1 million DAU and 100 million requests per day, cutting 1000 tokens from this description could save millions of dollars in inference costs per year.

Why Pi

Caleb apologizes for the negative tone, saying Anthropic has made a very cool piece of software and he does not think he would have done better if he had been in charge from the start. LLM coding harnesses are a new space, and later tools like Pi benefit from hindsight, he writes.

Still, he feels Claude Code lacks a strong vision or direction. It looks as if excited engineers and PMs are rapidly tacking on features that suit their own workflows without considering how they might make things worse for other users. There seems to be no one high up whose job is strictly to say ‘no’.

For now, he is leaving. The first few days with Pi required some finagling, but since then it has been smooth. He enjoys the stability, minimalist approach, and extreme customizability. He has already written a couple of extensions for his use cases and hopes to share them after cleanup. He encourages others feeling the same frustrations to give an open source harness a try.

END
Feedback