Claude Code is unusable now
TL;DR
Theo says Claude Code crossed from frustrating to unusable for his workflow — after paying $200/month and historically getting huge value from the plan, he hit a breaking point when Claude Code refused to help debug a simple Dropbox issue on his Mac and kept insisting it was only for software engineering tasks.
Anthropic appears to reject requests mentioning Open Claw at the API layer — Theo demonstrates that simply adding “personal assistant running inside of Open Claw” to the system prompt triggers a 400/API error, which he frames as an aggressive attempt to block people from routing Claude subscriptions through third-party harnesses.
The ugliest detail is the billing/routing behavior tied to system-prompt text — with extra usage disabled, the Open Claw mention gets rejected, but after Theo enables paid overage, the same request goes through, which critics like Simon Willison called “a really bad look.”
The real business conflict is GPU economics, not just policy wording — Theo argues Anthropic tolerated heavy Claude Code use because power users could consume up to roughly $5,000 of inference on a $200 plan while acting as marketing, but Open Claw users likely burned far more tokens without the same upside.
Even Anthropic-friendly developers say the rules are impossibly vague — Theo highlights Matt Pocock’s month-long struggle to get answers about whether wrappers, CI, agent SDK usage, or course materials are allowed, capped by Pocock’s line that Anthropic’s subscription rules are “more complicated than TypeScript generics.”
Codex became Theo’s replacement in real time — he shows Codex successfully researching Dropbox tray-icon issues, suggesting likely causes like duplicate installs, removing broken installs, and producing a reinstall checklist, which convinced him to repoint his terminal alias from
cc= Claude Code to Codex.
The Breakdown
Theo opens angry — and with a live failure demo
Theo starts by saying he didn’t even want to make another Anthropic video, but his audience pushed him there. Then he jumps straight into a demo: add a tiny system-prompt note saying Claude is a “personal assistant running inside of Open Claw,” ask “is Claude here?”, and get an API error. For him, that’s the tone-setter: this isn’t abstract drama anymore, it’s product behavior breaking in front of him.
Why Anthropic would want to shut down Open Claw usage
He lays out Anthropic’s economic logic pretty bluntly: the $200/month Claude Code plan can yield something like $5,000 of inference for heavy users, and that only works because average users underspend while power users create buzz and lock-in. Open Claw users, in his telling, are different — lots of idle heartbeats, bloated context, poor caching, and the kind of usage that torches tokens without helping Anthropic’s business. He even cites Mario from Pi noting that Boris, Claude Code’s creator, submitted PRs to Open Claw to improve caching and lower costs.
The ban moved from headers to system prompts
Theo says the first enforcement step made sense technically: reject requests based on headers or harness signatures tied to Open Code/Open Claw. But developers found a workaround by having Open Claw call Claude Code’s own CLI (claude -p) instead of using its native harness directly. Anthropic’s next move, according to Theo, was to also block references to Open Claw in the system prompt so even that wrapper path would fail.
The nastiest reveal: mention Open Claw and you get billed differently
This is the part that really set him off. With “extra usage” turned off, the Open Claw system-prompt mention causes a 400 error; with overage spending enabled, the request suddenly works. Theo calls that “dirty” because it suggests special routing or policy treatment based on system-prompt text, and he backs that up with reactions from Simon Willison and others who say this blows up the charitable explanation that Anthropic is only trying to manage caching and feature flags.
Matt Pocock becomes the human proof that nobody knows the rules
Theo reads Matt Pocock’s post almost as evidence in a trial. Pocock — whom Theo describes as a much nicer, smarter, more British version of himself — has a Claude Code course for sale and still can’t get a clear answer on what wrappers, SDK usage, or CI cases are allowed after weeks of trying. The killer joke lands because it’s true: Anthropic’s subscription rules are now “more complicated than TypeScript generics.”
The Dropbox incident that killed Claude Code for him
Theo explains the actual workflow he still relied on Claude Code for: front-end/UI work, setting up new machines, and debugging weird local computer problems. On a new laptop, Dropbox was stuck launching without showing a menu icon, so he asked Claude Code to kill and relaunch it. After one attempt, Claude started refusing to help, saying Dropbox UI behavior was outside its area instead of just using its search tools — and Theo’s reaction is basically, are you kidding me, that used to work.
His conspiracy theory weakens — but his conclusion doesn’t
Theo initially suspects Anthropic quietly changed Claude Code’s system prompt to narrow it to pure software-engineering tasks and stop people from using it like a general-purpose local assistant. Then Bad Logic Games, creator of Pi, chimes in with tracked prompt history suggesting there wasn’t a meaningful visible prompt change, which pushes Theo toward another possibility: extra system-prompt injection at the API level. Either way, he says the user experience has obviously changed for the worse.
Codex wins the head-to-head, so Theo changes his habits
He reruns the Dropbox problem in Codex and gets the behavior he wanted: process inspection, web research, recognition that this is a known recurring macOS Dropbox tray issue, a suspicion about duplicate installs via Homebrew, and finally a clean uninstall plus reinstall checklist. That’s enough for him to make it symbolic — he changes his cc alias in zsh from Claude Code to Codex with --yolo, says Claude Code changed how he built software back in December, but declares this the last day he’ll intentionally open it for real work. He closes by saying OpenAI’s team feels more transparent, more responsive, and more eager to win, while Anthropic keeps making things worse.