One Developer Left Claude Code. The Real Choice Was the Tool Around the AI
coding harness
Software that connects an AI model to tools and a coding workflow.
system prompt
Instructions that shape how the AI behaves.
AGENTS.md
A project instruction file read by coding tools.
What happened
The post Goodbye Claude, Hello Pi describes one developer's switch. The developer had used Claude Code for a long time. Anthropic, the company behind Claude, makes Claude Code, a tool for AI-assisted coding. Pi is an open-source coding harness that can be extended. Here, a harness means software that connects a model to files, terminals, instructions, and a developer's workflow.
The author retired Claude Code as their main harness. They moved their main setup to Pi. The post praises Claude Code's importance, but lists many small frustrations. It does not present the move as a controlled product test.
The problems were in the workflow
The author focused on Visual Studio Code integration. They said only one Claude Code session could receive selected text from the editor. That makes parallel work harder. One session might review code while another starts a different task.
The author also criticized Claude Code's official VS Code extension. Some slash commands did not work there. The extension lacked a custom status line. Output from several subagents became hard to follow in one conversation.
Instruction files created another source of friction. The author described maintaining CLAUDE.md and AGENTS.md separately. They said Anthropic announced AGENTS.md support one day after the decision to switch. That timing mattered because the author had already reorganized the workflow.
Fast updates can also create friction
The author argued that Claude Code added features faster than it improved stability. They counted 244 releases in 2026. They said more than three quarters changed the system prompt or tool descriptions. They also estimated that the default Bash tool description used nearly 3,000 tokens. These counts and estimates come from the author. The post does not provide an independent release audit.
This detail matters because changes to prompts and tool descriptions can alter how a coding agent behaves. The author wanted a smaller and more predictable setup. That is a personal judgment, not a measured finding about every Claude Code version.
Why the choice matters
The story is not mainly about one model beating another. It is about the software around a model. A coding harness controls how the tool receives context, connects to an editor, shows work, and accepts extensions. Those choices can shape a developer's daily experience.
The author described Pi as more minimal and customizable. After several days of adjustment, they said Pi felt smooth. They also wrote extensions for their own workflow, including support for multiple open tabs. The post does not report a benchmark, a cost study, or a code-quality comparison.
The switch also does not prove that the author stopped using Anthropic models. A person can change the harness while keeping some underlying services. The source describes a change in the working layer.
What is confirmed, and what is not
The confirmed event is narrow. One long-time user said they moved their primary coding setup from Claude Code to Pi. The post records the reasons they gave. It does not show a wider migration, a company response, or a general verdict on either tool.
The supplied Hacker News listing showed 367 points and 228 comments. Those figures measure community attention. They do not prove the author's claims. The comments themselves are separate from the article's evidence.
We still do not know how common the reported IDE problems are. We do not know whether Pi remains better for the author over time. We also do not know whether the promised extensions will be published, or whether later Claude Code updates address the complaints.
What to watch next
The useful next evidence will be practical. Readers can watch for the author's extensions, Claude Code's multi-session behavior, and its AGENTS.md support. A stronger comparison would use the same tasks and models in both tools. It would record setup time, cost, reliability, and the effort needed to review changes.
Sources: caleb.software and Hacker News.
Why One Developer Switched Coding Helpers
📰 Full story: One Developer Left Claude Code. The Real Choice Was the Tool Around the AI
One developer moved from Claude Code, a coding helper, to Pi, an extendable open-source tool.
instruction file
A file that tells the coding helper how to work.
subagent
A smaller helper working inside the main coding tool.
Hacker News
A website where people discuss technology news.
💡 The gist
- One developer moved their main coding setup from Claude Code to Pi.
- Small workflow problems pushed the decision.
- Hacker News attention did not prove the story was correct.
Anthropic is the company behind Claude. Claude Code is Anthropic's coding helper. Pi is an open-source coding helper that people can extend.
The writer used Claude Code for a long time. They still respected the tool. They said many small problems had piled up.
In VS Code, only one connected Claude Code session could receive selected text. That made two tasks harder to run together. A developer might want one session to review code. They might want another session to begin new work.
The official VS Code extension also caused trouble. Some slash commands did not work. It had no custom status line. Several subagent replies became hard to read together.
The writer also managed two instruction files. They were CLAUDE.md and AGENTS.md. These files tell a coding helper how a project should work. The writer said AGENTS.md support arrived one day after switching. The update came too late for their decision.
The writer counted 244 Claude Code releases in 2026. They said more than three quarters changed prompts or tool descriptions. They estimated almost 3,000 tokens in the default Bash instructions. Those numbers came from the writer. They were not an independent audit.
The deeper lesson is about the tool around the model. A coding helper must receive context. It must work with an editor. It must show results clearly. It must support the user's habits. A smarter model cannot fix every workflow problem.
The writer said Pi became smooth after several days. They also made extensions for their own work. The post does not show that Pi is faster or cheaper. It does not compare code quality.
This is one person's story. We do not know if many users share these problems. We also do not know whether the writer still prefers Pi later. The post does not prove they stopped using Anthropic models.
The supplied Hacker News listing showed 367 points and 228 comments. These numbers show attention. They do not prove the article's claims. The comments are not evidence presented by the original post.
Next, readers can watch Pi's extensions. They can watch Claude Code's editor support. A fair test would use the same tasks and models. It would measure time, cost, and repair effort.
💬 HN comment digest: Pi is flexible, but it takes more work
People who like Pi value its simple base and ability to change it. Other comments point to problems with long conversations and service restrictions. The speed and bug reports are individual experiences.
- Pi is more like a base that users extend. Claude Code is more like a ready-made tool with more pieces already connected.
- One user reported about 150 tokens per second on an M1 Mac with Pi and kapitsa. That is not a promise for every computer or model.
- When a long conversation is compressed, Pi may lose track of the next step, according to one user. Some users report better results with pi-vcc or handoff notes.
- Comments say Codex is easier to use from other harnesses, while Anthropic access has had tighter limits and extra technical details. Claude now supports AGENTS.md, but .agents/skills may still be missing.
- GUI and TUI each have advantages: GUI is better for history and remote/mobile use, while TUI is better for terminal control. The better choice depends on the user’s workflow, model, and need for control.
growing digest at 406 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
A Programmer Chose a New AI Helper
📰 Full story: One Developer Left Claude Code. The Real Choice Was the Tool Around the AI
One person changed from Claude Code, a coding helper, to Pi, another helper, after many small annoyances.
Hacker News
A website where people talk about technology.
Pi
A coding helper that people can change.
Anthropic makes Claude, an AI that helps with computer programs. Claude Code is a helper for making code.
The person used Claude Code for a long time.
The tool sometimes made work harder.
Only one work session could receive selected words from the computer editor.
The person wanted two work sessions.
That was difficult.
Some commands did not work in another screen.
Many answers also became messy to read.
The person chose Pi because Pi could be changed.
They added small tools for their own work.
This is one person's story.
We do not know if everyone would choose Pi.
We do not know if Pi makes better code.
Hacker News is a website where people discuss technology.
The story had 367 points and 228 comments there.
That means many people noticed it.
It does not mean every claim is true.
People can watch both tools change.
💬 HN comment digest: Pi is a box you can rebuild
Some people like Pi because they can change it. Some people like Claude Code because it is more ready to use.
- Pi is like a small toolbox. You can add new parts. Claude Code is like a tool that already comes assembled.
- One person said Pi was very fast on an M1 Mac. Another said Pi sometimes forgot what to do after a very long chat was squeezed smaller. Some add a fix or keep a handoff note.
- Codex and Claude have different rules for using them in other tools. AGENTS.md works in Claude Code, but some skill folders may not. Some people like GUI screens, and others like terminal screens.
growing digest at 406 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 HN comment digest: Pi trades turnkey behavior for control
The comments present Pi as lightweight, customizable, and easier to observe, while raising real concerns about reliability and Anthropic integration. Performance and failure reports are user anecdotes, not results from a common benchmark.
growing digest at 406 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.