I’ve been a longtime Claude Code user, always since it became accessible in the Claude Pro subscription scheme complete a twelvemonth ago. I recall the archetypal period I used it, emotion giddy as it spat out messy Python code for any random weekend project I was operating on, all from a sole prompt. I’m certain that in a few years we’ll all appearance rear at its publish as among the important drivers of this new era of application betterment we’re in.
Despite this, because of an gathering of many small frustrations, I’ve eventually decided to quit it as my chief coding harness and migrate my betterment setup to use Pi instead.
Someone at Anthropic must’ve caught wind… fair a day following making my decision, they announced Claude Code would finally commencement supporting AGENTS.md records in an attempt to get me to stay. Unfortunately, for me it was too little, too late.
Complaining
In lieu of a appropriate obituary, I’d akin to obtain this period to say a few words concerning the things I liked smallest concerning Claude Code (and thus what I’m looking onward to avoiding alongside Pi).
The AGENTS.md item benevolent of epitomizes the entire logic I decided to discontinue using Claude Code. Don’t get me wrong, having a ton of CLAUDE.md records scattered concerning assorted GitHub hobby repos is awesome guerrilla marketing. But it sucks for users to have to agreement alongside copying or symlinking between the two formats.
Like the Luddite I am, I motionless favor to assessment most of the code it churns out before I commit it. I use(d) Claude Code inner VSCode, and it lets you nexus the harness to the IDE. This is nice since you can emphasize content in a document and have it automatically included in the environment whenever asking questions or requesting changes during a review.
Except you can lone have one meeting “connected” to the IDE at a time. If you open a second tab to commencement another project during the archetypal one works, it can’t obtain the content choice events from the IDE at all. This has been reported at least as far rear as December 2025… and closed as a copy of a totally unrelated bug.1
Of course, you can continually use the authoritative Claude Code VSCode GUI extension, which does authorize you to have multiple tabs open and burden the content choice into the environment of each. But this expansion has plentifulness of problems of its own. Half of the built-in slash commands don’t activity at all in the GUI. It doesn’t have any assistance for a tradition position line. Subagent output messages are merely layered into the chief conversation thread in chronological order, which is a disorder to peruse whenever additional than one is running. I could go on and on.
In spite of all of these existing issues, it seems akin Anthropic has worked at a feverish gait to group additional features into Claude Code without much attention for norm or stability. Dynamic workflows, agent teams, ultrareviews – all awesome ways to assistance endeavor customers scorch always additional tokens all day, but not really item I desire or need.
Alongside all of these characteristic releases, they continually appear to be fiddling alongside things in ways that disorder up my individual setup. At the period of writing, there have been 244 releases to Claude Code this year; complete 3/4 of these changed the scheme immediate or tool descriptions, which can substantially alter how the LLM performs (especially whenever these new instructions refuse person prompts or repo conventions).
And what are all those changes for? Despite chopping 80% of the scheme prompt, Claude Code motionless has way additional overhead than Pi for apparently no benefit. The default Bash tool’s clarification is nearly 3,000 tokens, most of which is random material concerning using git and beginning PRs on GitHub.2 The models already cognize how to do this without detailed instructions, by the way.3
Looking Forward
This article has been a lot of hating. I’m sorry! That’s not my customary style, and I do think Anthropic has made a really chill part of application here. I’m additionally not under any misconception that I would’ve done a improved job if I were the one in accusation of operating Claude Code from the beginning. LLM coding harnesses are a extremely new area and being among the archetypal to innovate in a site is a really difficult task. Later tools akin Pi have the advantage of hindsight to see what plant and what doesn’t, what’s a center characteristic and what can be cut.
Ultimately, CC feels akin a merchandise without a powerful imagination or direction. It’s as if a bunch of enthusiastic engineers and PMs are quickly tacking on any features lawsuit their workflow… but without deliberation for how this power create things worse for another users. It seems akin there isn’t person high-up whose job is strictly to say “no” to these requests, which is incredibly crucial whenever it comes to merchandise engineering. So, for now, I’m leaving.
I’m enthusiastic to lean additional into Pi. The archetypal few days required several finagling, but since afterward it’s been gentle sailing. I’m enjoying the stability, minimalist approach, and the extreme flat of customizability. I’ve already written a brace of extensions to assistance my use cases, which I’m expecting to portion shortly following I’ve cleaned them up a bit. If you’re additionally emotion these frustrations, I’d extremely advance giving an open origin harness a try!
