Summary of "Claude Code is unusable now"

High-level claim

The presenter says Claude Code (Anthropic’s developer-focused Claude subscription) has become effectively unusable for his workflows due to recent API and policy changes from Anthropic — primarily changes that affect system prompts and request headers.

Technical issues and behavior changes

GPU / capacity pressure Anthropic is reducing GPU usage and tightening how subscription plans may be used (to preserve capacity for research/enterprise). This has produced limits and request rejections for certain usage patterns.
Header- and prompt-based blocking The API began rejecting requests when request headers or system prompts reference a third‑party agent/harness (the transcript uses the name “Open Claw” for that harness). If headers mention the harness, the API can return errors; mentioning it in the system prompt can also cause failures.
System-prompt routing / injection The presenter demonstrates Claude behaving differently depending on system prompt content and claims Anthropic may be injecting or modifying system prompts on the API side. That can change model behavior (for example, refusing requests outside software‑engineering help) and may trigger different routing or billing decisions.
Extra-usage flag oddity Toggling the account “extra usage” option altered whether requests with blocked system prompts were honored, implying Anthropic applies special routing/billing when certain system prompts are present.
Caching / rate-limit tradeoffs Anthropic appears to have chosen bluntly banning some harness usage rather than implementing finer-grained rate limits or caching controls that might preserve value for paid subscribers while blocking inefficient third‑party clients.

Practical impact and examples

Typical uses affected The presenter used Claude Code for:
- front-end UI work
- setting up/configuring new machines
- quick local debugging (non-code issues, e.g., fixing Dropbox UI behavior)
Demonstrated failure In one demo, Claude refused to assist with a Dropbox debugging task (saying it was outside its scope), while an alternative CLI (Codex) successfully ran commands, researched, and fixed the problem.
Workarounds attempted
- Programmatic proxy via the official Claude CLI (previously used to bypass header bans), but this still required injecting the harness’s system prompt and was blocked.
- Enabling “extra usage” to force a different routing/billing path, which temporarily changed behavior.
Result Due to unpredictability and opaque policy, the presenter removed his quick-access alias for Claude Code and switched to a different tool/CLI (Codex) for local tasks.

Product features, SDKs, and tooling mentioned

Claude Code subscription Claude for software engineering/agent use — includes an Agent SDK, a Claude CLI, and special token/subscription billing models.
Agent SDK / harnesses Third‑party harnesses (the “Open Claw” harness in the transcript) that call the Anthropic API and may have different caching/heartbeat implementations; these inefficient harnesses are a focal point of the friction.
Claude CLI behavior Used both interactively and as a programmatic proxy; subject to header/prompt-based blocking.
Extra-usage (billing toggle) An Anthropic dashboard option that changes how overages are handled and appears to affect routing when blocked prompts exist.
Codex An open-source CLI/model used by the presenter as an alternative for coding, research, and local tasks. Described as more transparent and easier to build on.
T3 Code / T3 Canvas / T3 Chat The presenter’s open-source tooling that integrates Claude/Codex via the Agent SDK or local CLI.
Clerk (sponsor) Developer identity and billing components used in the demo — noted for easy integration of auth, organizations, and billing features.

Policy and community reaction / analysis

Lack of clarity Developers (for example Matt Pocock) complain about vague and inconsistent rules for subscription usage — unclear distinctions between personal vs commercial vs CI vs open‑source/redistributable cases.
Community criticism Several developers (Simon Willison, Dex, others) criticized treating requests/billing differently based on system-prompt text, calling the approach heavy‑handed and opaque. Some defended the policy as a response to inefficient third‑party harnesses causing cost spikes, but many want clearer rules.
Mixed evidence on prompt changes One community member (Bad Logic Games) tracks Claude Code system prompts and reports no meaningful change over time, which complicates the presenter’s theory. The presenter still suspects API-side injection or routing-level changes.

Advice and actions taken by the presenter

Switching defaults The presenter replaced his Claude Code CLI alias with a Codex CLI alias for quick local tasks.
Recommended alternatives Try Codex or other tooling if you rely on local/CLI agent workflows. Prefer transparent, open-source CLIs/harnesses (Codex, T3 Code, Pi) or different auth/harness approaches for more predictable behavior.

Summarized recommendations

If you rely on Claude Code for local debugging or agent-like tasks, be prepared for disruptions and consider alternative CLIs/models such as Codex.
Watch Anthropic’s policy updates and demand clearer documentation on subscription usage, system-prompt handling, and programmatic access.

Main speakers and sources referenced

Video presenter (unnamed in transcript) — primary speaker and tester of Claude Code
Anthropic (Claude team; Boris and Thorik referenced as Anthropic staffers)
Matt Pocock — developer who criticized the lack of clarity and created a Claude Code course
Mario / Bad Logic Games — creator of the Pi agent; tracks system prompts and contributed analysis
Simon Willison and “Dex” — community figures who criticized the prompt-based billing/behavior changes
Codex — alternative CLI/model used and recommended by the presenter
Clerk — sponsor and product highlighted in the video