Claude Code is unusable now

Anthropic has implemented a new enforcement mechanism that scans system prompts for banned phrases like «OpenClaw» and rejects requests outright—even for paying subscribers. The implementation is even more cynical than a simple block: if you have «extra usage» enabled in your account settings, the same request that would fail with a 400 error suddenly succeeds, but you're billed at a higher rate. This means Anthropic has built routing logic that deliberately treats identical API calls differently based on a hidden flag tied to system prompt content.

The rationale appears to be resource conservation: OpenClaw users burn tokens inefficiently due to implementation issues like excessive context in heartbeat pings and poor caching. Boris, Claude Code's creator, even contributed PRs to OpenClaw to fix caching and reduce costs. But instead of better rate limiting or technical solutions, Anthropic chose a blanket ban enforced through opaque prompt scanning. The result is that users who mention any third-party tool—even in personal assistants—are flagged, blocked, or secretly upcharged.

This goes beyond terms-of-service enforcement. It's hidden behavior modification that users discovered only through trial and error, with no official documentation or warning beyond vague emails. Even developers who want to comply have no way to know what phrases will trigger the filter or how the billing logic actually works.

The host's theory: Anthropic secretly modified Claude Code's system prompt to make it refuse non-engineering tasks, ensuring that the `claude-p` command—used by OpenClaw as a workaround—would stop being useful for general assistance. The evidence is circumstantial but compelling: yesterday, Claude Code happily debugged system-level issues; today, it flatly refuses, saying «that's outside my area.» The phrase «software engineering tasks» appears repeatedly in rejections, suggesting new prompt-level constraints.

Bad Logic Games, creator of the Pi agent, tracks Claude Code's system prompt over time and reported no meaningful public changes—but the host notes that Anthropic injects additional instructions at the API level that wouldn't appear in the client. Until Anthropic releases full receipts, the suspicion remains: they're narrowing Claude Code's scope to make third-party integrations less viable, all while refusing to document what's actually allowed. The lack of transparency has turned even defenders into skeptics.

Two days before this video, Boris (Claude Code's creator) replied «Yep» to a question about using subscriptions with personal local tools that wrap Claude harnesses via the agent SDK. The host was ready to publicly defend Anthropic and declare the drama over. Then Thoric, Anthropic's communications lead, replied to clarify: «To be clear, this is not guidance or an update on the agent SDK. We're still working on clarity there.» In other words, even the creator's informal thumbs-up couldn't be trusted as policy.

This incident crystallized the problem: Anthropic has no coherent answer to give, and every attempt at clarification creates more confusion. Developers are left in limbo, unable to ship products, teach courses, or even use the tools they're paying $200/month for. Meanwhile, competitors like OpenAI—despite being worth over $800 billion—are providing clear terms, responsive support, and open-source tooling. The contrast has become impossible to ignore.