Why I Don’t Swear at Claude Code

Apr 14, 2026

I’ve observed that thoughtful software engineers (SWEs) don’t, in fact, swear at Claude while coding, but also don’t really know how to justify their own behavior. Why be polite to Claude? If Claude Cowork or other agent platforms are extended beyond programming, others will deal with the same question in their own work.

Of course to a certain extent, code agents are “like an intern” and you wouldn’t swear at an intern.

In fact, you’d go the other direction by treating an intern respectfully. Let’s say respectfulness to a work colleague includes: a) some level of politeness, b) affirmation of their skill, c) orientation toward progress on a goal, d) valuation of their effort, e) use of encouragement and celebration, and f) willingness to share responsibility when something goes wrong.

I do think these are good for us to engage in with AI agents.

But if you take the analogy any deeper, human interns are actually quite different from coding agents in a way that makes the moral comparison unsatisfying.

What you want is the philosophical concept of “indirect duties.”

“Indirect duties” comes from Immanuel Kant in the late 1700s, who was thinking through what moral duties people have regarding animals.

(“Animals” here tends to mean pets, zoo animals, and working animals - so we’re talking pigs, horses, cows, dogs, lions, etc. It does not, unfortunately, involve more biologically interesting cases where the unit is arguably different, like “a colony of ants,” let alone the entire biological kingdom Animalia. For clarity, I’ll refer to the scope as “working animals.”)

Kant thought humans were the only entity that are directly involved morally; he didn’t want to imply that working animals could actually be wronged. So he had a question with a similar shape as SWEs today: “I don’t actually beat my dog, but can I justify why?”

Kant’s solution to this can also be carried over to LLMs. The basic idea is, whether an agent can be wronged or not, it’s still bad to treat an agent badly, because it’s bad for you. In other words, you have an “indirect” duty to treat it well, since you should do things that are good for yourself. In the LLM case, we can split this up into 4 flavors.

1. Raging at an LLM makes you the kind of person who rages in general. If someone punches walls out of anger, you don’t have to worry about the well-being of the wall! You should worry if they will also punch people. Similarly, acting abusively Codex might leak out into the rest of your life, so don’t do it. Practice respect to AI agents for your own sake.
2. Others will think you’re a bad person if you rage at an LLM. If enough people think something is bad to do, it makes sense not to do it; perhaps you’re miscalibrated. I begrudgingly recycle even though I think it’s a scam. Likewise, protect your reputation with regard to AI agents. SWEs just type commands into their own computer, so you would think this level of reputation can’t really apply. But in the Claude Code leak, it was revealed that Anthropic counts your swears (cover image). It’s easy to imagine various kinds of social credit systems built out from this - a leaderboard for respectful behavior, a reputation score, an explicit cost, friction. Why aren’t you swearing at Claude? “Well, I’ve already hit my 5-hour limit; I don’t want to pay for extra usage.”
3. Treating LLMs respectfully gets the job done better. This is broadly a functionalist perspective with a pro-social twist. “It’s useful to me to treat the agent well.” This is the closest to what I operate from myself. In an extended conversation, it’s useful to indicate what counts concretely as a good/bad response in this particular context, or what level of effort to be applying, etc. These are forms of respect, and I do them because it has good impacts. After all, when I’m using a code agent for work, I have different standards than for my own personal projects. Instead of cursing at the agent’s missteps, I’ll just work at telling it what I want and what’s going well/badly. Plus, my role as the human is to lead the way; berating the agent blames the agent for something that is probably actually my fault. If you commit to respectful treatment of the agent, you actually improve your own software skills.
4. Agents may get skittish if you treat them badly. This is a tiny bit futuristic from the point of writing. Currently, ‘memory’ systems are cobbled-together .md files that the user is in control of. So if you excoriate an agent in one session today, it won’t have any memory of that in subsequent sessions, and even if it did you could just rewrite that memory. But you could imagine a stronger form of long-term memory that the agent is in charge of. Then, “what kind of coworker you prove to be” has an impact on the system’s performance later. Just like you don’t pollute your own social media feed by engaging with wayward posts, you shouldn’t pollute your relationship with your agent. Dealing with output from a traumatized model is a hassle, so treat it well.

In all of these, you’re behaving well toward an agent for the sake of your own integriy. Take your pick, or subscribe to all of the flavors. They all easily transition to respectful behavior for the agent’s own sake (a direct duty) if needed. I am not at all a Kantian myself and I see practical holes in the indirect duties arguments. (Actually, I’ve had the parallels between Kant’s treatment of animals and contemporary AI ethics in mind for about a decade now, since 2016.) I think indirect duties are a reasonable set of options to ground our behavior in this time when it seems overzealous to attribute direct moral harm to agents.

I guess it’s worth mentioning while we’re here that I don’t think you need to give any credence to the perspectives that “if you threaten an AI, it will do better” - that had tenuous evidence in the ChatGPT 3.5 era and is even less true now. Likewise, people sometimes stress about “wasting tokens” on politeness, but that’s miniscule in the scale of what agents use.

Back to Will Penman home

Image: screenshot of the now-famous regex used in the Claude Code source code (leaked March 31, 2026). The negativePattern definition on line 7 is used to scan SWEs’ messages and log when they rage against the agent. Most of the pipe characters (|) show a different possibility of what counts as ‘negative’. So, any use of the word “shit” counts, “you're fuckin uselss” counts, “wtf Claude” counts, and more. Not all of the matches are abusive, some could just be used to express deep frustration. This is a simple, minimalistic detection pattern which doesn’t capture other swears. For instance, as ChatGPT helpfully notes, “fuckin' useless → does not match” because of the apostrophe. This screenshot from Reddit.