Claude Code: 244 releases in a year, 3000 tokens in Bash — and a developer moved to Pi

⚡ The gist in 5 seconds

• The author of the caleb.software blog abandoned Claude Code in favor of Pi after a year of working with it.
• A day after his decision, Anthropic announced AGENTS.md support in Claude Code.
• Over the year, 244 releases shipped, more than 3/4 of which changed the system prompt or tool descriptions.

🔍 What happened

The post "Goodbye Claude, hello Pi" on caleb.software racked up 681 points and 424 comments on Hacker News — a lot for a personal blog post. The author writes that he had been on Claude Code since it appeared in the Claude Pro subscription over a year ago. He's leaving not because of a single bug: small irritations piled up. A day after his decision, Anthropic announced that Claude Code would start reading AGENTS.md files. The author himself calls the coincidence funny and adds: it was too late.

The mechanics of the irritation are concrete. Context files had to be copied or symlinked between CLAUDE.md and AGENTS.md. In the VSCode integration, only one session can be attached to the IDE — open a second tab, and text selected in the editor no longer lands in it. The bug had been open since December 2025 and was closed as a duplicate of an unrelated issue. The official GUI extension supports multiple tabs, but half of the slash commands don't work in it, there's no custom status line, and subagent output is dumped into the shared chat in chronological order.

The pace takes separate flak. Over the year, 244 releases of Claude Code shipped, more than 3/4 of which touched the system prompt or tool descriptions — that is, they changed the model's behavior. The description of the standard Bash tool pulls almost 3000 tokens, and most of it is about git and GitHub pull requests. Anthropic cut 80% of the system prompt, but the overhead is still higher than Pi's. By "enterprise features" the author means dynamic workflows, agent teams and ultrareviews — in his words, ways to burn more tokens per day.

🧠 Our take

Our assessment: the post took off not because someone left a tool, but because the departure is described with numbers. "It's inconvenient for me" doesn't get read on Hacker News, but "244 releases, 3/4 changed the prompt, 3000 tokens on the Bash description" — that gets read, and people recognize their own experience in it. The argument here isn't about taste, but about what counts as stability in a tool that affects your code every single day.

We believe the problem is more real than it seems from the comments. A harness isn't just the model; it's also the layer of instructions on top of it. If that layer is edited several times a week, the user's agent behavior changes without their involvement, and reproducibility breaks down. In this logic, AGENTS.md isn't a gift but a belated admission that the external format won. You'll still have to copy files between the two names until auto-detection appears.

Who wins? Pi gets free marketing from 681 points and 424 comments — you couldn't invent a better channel than someone else's frustration. Anthropic isn't losing the enterprise (they pay for those very agent teams), but precisely the evangelists — the people who carry the tool into their teams. The loser, in our assessment, is the one paying in tokens: bloated tool descriptions eat up context before the model has even read your task.

⚠️ What we don't know

The source says almost nothing about Pi itself: no license, no pricing, no benchmarks, no community size — only "a late tool with the advantage of hindsight." The figures about 244 releases and 3000 tokens are one person's counts, not an audit. The link between the AGENTS.md announcement and the post rests on a coincidence of dates, and the phrase "someone at Anthropic must've caught wind" is the author's joke, not a confirmation. Worth checking yourself: open the Claude Code changelog and see how often tool descriptions changed there.

Prediction: within a quarter, Claude Code will get automatic AGENTS.md pickup without manual symlinks, but there will be no public promise to freeze system prompt edits — the release pace will continue. Verifiable via their changelog.

Disclaimer: a single source — a developer's personal blog; the figures and bug descriptions are not officially confirmed.

Source: caleb.software