Anthropic bans cruelty to Claude, and a developer builds a profanity filter for prompts

On October 8, Anthropic updated the usage rules for its models. One of the changes looks odd for a user agreement: "systematic senseless cruelty" toward models is now banned. A Habr author who writes about Claude Code noticed it while preparing a news item. His own skill, Prompt Warrior, analyzes his logs of conversations with the agent, and it shows profanity in almost every other message. He drew an engineering conclusion: he did not change his habits, and instead wrote a tool that rewrites his prompts automatically. The result is a kind of Punto Switcher for requests to AI.

What changed in the rules

Going by the headline of the TechCrunch piece, the update touched two areas: model abuse and interference in elections. The source does not include the full text of the policy, so the criteria for "systematic" and "senseless" remain at the level of the wording. But the logic is readable. The rule is not about punishing an irritated remark when the agent writes "done" for the tenth time on top of broken code. It is about repeated abuse with no task behind it.

This continues a line Anthropic has followed for some time. The company does model welfare research, and in August 2025 it gave some versions of Claude the ability to end a conversation when a user's behavior is persistently abusive. Now the same approach is written into a legal document.

How the rewriter works

The author starts from a practical problem. He does not swear in order to abuse the model: it is an emotional reaction to an agent that reports success without checking. Many people who work with coding agents every day behave the same way. The idea behind the solution is simple: a layer between the human and the model that softens the wording before sending, the way a keyboard-layout switcher fixes text on the fly.

What is known from the announcement:

• date of Anthropic's rule change: 8 October 2026;
• subject of the ban: systematic senseless cruelty toward models;
• statistics tool: the Prompt Warrior skill, which analyzes logs of conversations with the agent;
• measurement result: profanity in almost every other message;
• target agent: Claude Code;
• approach: automatic text replacement instead of a change in behavior.

The available excerpt does not disclose implementation details. It is unclear whether the tool works with rules, a dictionary or a separate model, and how accurately it preserves meaning. The quality can only be judged from the full article.

Such a scheme has an obvious downside. Emotional coloring in a prompt carries a signal: "you did the wrong thing, and it matters." If the filter flattens a remark into a neutral one, the agent may understand less well that the error is critical. A good rewriter should remove the swearing but keep the demand.

How to avoid trouble

The practical takeaway for people who work with agents a lot is more modest than it may seem.

First, a single rude remark, judging by the wording, is not what the rule targets. The ban describes a system and senselessness. There is nothing alarming about a one-off "what the hell are you doing."

Second, if you swear in every other message, that is a reason to look at your process, not only at your vocabulary. A common cause of irritation is an agent that closes a task without checking. The cure is not a filter but a requirement in the project instructions: run the tests and show the output before saying "done."

Third, a client-side filter makes sense as hygiene. It is cheap and does not break the workflow. But you cannot rely on it as a guarantee of policy compliance: the company has not yet explained in detail how exactly Anthropic will detect violations and what counts as systematic.

One interesting side effect: there is now a measurable metric for how we talk to AI. If you already have logs, you can count your share of swearing in an evening. The result may surprise you.