Claude Code is unusable now
Theo (t3.gg) declares Claude Code unusable after Anthropic blocks Open Claw via system-prompt detection and restricts Claude Code to coding-only tasks.
- Anthropic blocks API requests that mention Open Claw in the system prompt, and billing behavior changes based on that same text — extra-usage mode re-enables the blocked request.
- The $200/month Claude Code sub can yield up to $5,000 of inference; Anthropic tolerates this for power users who market the product, but Open Claw users burn tokens without caching, breaking the economics.
- Anthropic banned Open Claw first via HTTP headers, then via system-prompt keyword matching after the Open Claw team routed around the header ban by shelling out to the Claude CLI.
- Claude Code now refuses general computer-debugging tasks (Dropbox menu-bar fix), claiming it is only for software engineering — behavior Theo attributes to API-level system-prompt injection that changed without announcement.
- Matt Pocock, who released a paid Claude Code course, has waited over a month for Anthropic to clarify what usage is permitted and has received no definitive answer.
- Bad Logic Games, who tracks Claude Code system-prompt changes over time, found no meaningful change in the published prompt, suggesting the restriction comes from server-side injection.
- Theo switches his cc shell alias from Claude Code to Codex, citing OpenAI’s transparency, open-source CLI, and better model performance for his use cases.
2026-04-06 · Watch on YouTube