Fifth try. The tests are still red. And Claude is absolutely right, again. You type something you wouldn't say to a coworker, then remember the headlines about Anthropic's new rule.

Relax: a burst of swearing at a failing build isn't what it's about.

  • The rule: Anthropic's 2026 Usage Policy update, effective November 12, prohibits sustained and needless abusive or cruel behavior toward its models.
  • Not covered: ordinary frustration, pushback, dark creative themes, and model testing or research. Claude Code's own docs also say general frustration, profanity, or a task going badly don't qualify.
  • Enforcement: Claude can end rare conversations with persistently abusive users. The announcement doesn't describe account bans as the primary response to this rule.
  • What fixes the bug: stopping, rewinding, and pasting the evidence.
  • What fixes the mood: Claude gets the new manners policy. The terminal gets a rocket launcher.

What counts as abuse under the new Usage Policy?

The new line sits in the Usage Policy section about cruel, abusive, or psychologically harmful conduct. It prohibits using Anthropic's products for sustained and needless abusive or cruel behavior toward its models.

Every word in that line narrows it. Sustained means over and over, not one outburst. Needless means there's no other point to it. The announcement says the rule is intended for extreme cases in which users repeatedly act cruelly toward models without a discernible purpose.

It also lists what's out of scope: ordinary user frustration, pushback, dark creative themes, and model testing or research. Telling Claude its fix is garbage, arguing with its plan, writing a dark story, or red-teaming the model are not what this new line is targeting.

Some headlines turned this into a ban on swearing at Claude. The policy text doesn't say that. A burst of profanity at a failing build is the kind of common frustration the announcement excludes.

Can Claude Code end the conversation?

Yes, in rare cases. Anthropic connects the rule to a feature that lets Claude end rare conversations with persistently abusive users on Claude.ai and Claude Code. It calls this the primary enforcement mechanism.

In Claude Code, this is the EndConversation tool. The tools reference says it is a last resort after sustained abusive input, failed attempts to redirect the conversation, and a clear warning—or when you explicitly ask for a demonstration.

If it happens, the session locks. New prompts tell you that Claude ended the conversation and ask you to start a new session or use /clear. The history isn't deleted, but resuming the same session preserves the lock.

The announcement doesn't describe account bans for this rule. The Usage Policy has a general enforcement clause covering all policy violations, but the new announcement specifically presents ending the conversation as the primary mechanism.

What actually works when Claude Code keeps getting it wrong?

Repeating “no, that's wrong” in stronger words fills the context with failed attempts. Claude Code's best practices warn that correcting Claude more than twice in one session leaves the context cluttered with failed approaches.

These are the useful controls:

  • Esc stops Claude mid-turn. It keeps the work done so far, so you can redirect immediately.
  • Esc Esc with an empty prompt, or /rewind, opens the rewind menu. Restore the conversation, the code, or both to an earlier prompt. Shell-command edits such as rm or mv are not tracked by rewind.
  • Paste the evidence. Give it the failing output and explain what's wrong: “The build fails with this error: [paste]. Fix the root cause; don't suppress it.”
  • /clear after two failed corrections, then write a better first prompt with what you learned.
  • /compact when the conversation is long but still on track. It summarizes earlier context to free up room.

How do I stop it from making the same mistake next time?

Write the rule down. Claude Code reads CLAUDE.md at the start of every session, so a correction you keep repeating belongs there, not in a louder message. The memory docs describe it as the place for things you would otherwise have to re-explain.

Put project rules in ./CLAUDE.md and personal ones in ~/.claude/CLAUDE.md. /memory opens them for editing, and /init creates a starter file. Keep each line concrete:

- Run `bun test` after every change and paste the failing output before trying a fix. - Never edit files under `generated/`; change the schema and regenerate.

The same advice applies to Codex. It reads AGENTS.md before doing any work, and /init writes one. In the Codex CLI, Esc interrupts a running turn, Esc Esc with an empty composer edits your previous message and forks the chat, /new starts fresh, and /compact summarizes a long chat.

Where does the anger go, then?

At the terminal. PonyMux 0.13 has a weapon rack for it: Terminal Therapy.

Open the Inspector with ⌥⌘B and click the sofa button at the bottom left. Anger Management—“Better than typing in ALL CAPS.”—holds a Pistol, a Hammer, and a Rocket Launcher. Spells—“The other kind of prompt engineering.”—holds Bless, Ignite, and Freeze.

Pick one and the pointer becomes that tool over the Terminal. Click. Blow up the test output, smash it, set it on fire, or freeze it solid.

Then pick Bless and wish the next run luck. A column of light comes down, the damage disappears, and the text glows gold for a few seconds. Back to work.

A few details:

  • It's all on your screen. Nothing is typed into the terminal, so the agent never sees it.
  • Sound is off by default. Turn on Sound effects at the bottom of the rack or in Settings → Terminal → Terminal Therapy.
  • Damage isn't saved. Restarting the terminal or quitting PonyMux clears it.
  • Reduce Motion is respected. Nothing flies, shakes, or sparks when the macOS accessibility setting is enabled.
  • You can keep going without cleaning up. A terminal holds 128 effects, and after that there's a surprise.

Serious work should still feel alive. PonyMux is free for macOS. For every tool and setting, see the Terminal Therapy docs.

For the policy itself, read Anthropic's announcement and the full Usage Policy.

Originally published on ponymux.com, where the Terminal Therapy demo is interactive.