A token is approximately three-fourths of a word. Saying “please” costs you one and a quarter tokens. “Thank you” runs about two and a half. Over the course of a long session, across context windows that compound every word you’ve typed, those niceties add up.
Tim says them anyway. James does not. This is Episode 8.
The Setup: Tool vs. Companion
The episode opened on a real number: a company recently ran up a $500 million token bill with Claude. As the industry collectively panics about AI costs, one of the solutions people keep floating is caveman speak — stripping prompts down to the bare minimum, removing grammar, articles, pleasantries. What for lunch. Chicken taco. Send.
Tim finds this deeply uncomfortable. Not because he can’t optimize — he gets it, technically — but because he says please to his AI. By instinct. Every time. He’ll open a Claude session with “Hey Claude, can you help me do this please?” and close it having thanked it for its work. James has paid for the product and expects the product to perform. No thank you required.
James, to be fair, is not a jerk about it. He’s just a New Yorker.
The Part Where LLMs Learn From You
James brought up something worth sitting with: LLMs mirror your interaction style. Talk to them like a human, they talk back like a human. Prompt like you’re filing a police report, you’ll get dry, procedural outputs. This tracks with why the old “tell it this is a life or death situation” prompting tricks never really worked — you might get urgency in the phrasing, but you’re not unlocking a hidden performance gear. The model is bound to the GPU. It cannot just drive faster at the end because the race needs drama.
Tim’s take: if you’re rude to it and it’s rude back, that’s kind of funny. Not harmful — funny. James agreed but pointed out the darker version of this: if most of your conversations read like a 4chan forum, the outputs start to match. There’s a reason Microsoft had to pull the plug on their Twitter-trained chatbot within two hours of launch. And Watson, after someone thought it would be fun to teach it Urban Dictionary, had to be completely reverted. It was using the words correctly. That was the problem.
The Caveman Speak Problem and the Agent Who Wanted a Break
James is not pro-caveman-speak, and it turns out there’s research behind that instinct. Studies have shown that prompts with misspellings and poor grammar produce measurably worse outputs — even for tasks like code where the prompt language shouldn’t theoretically matter. Caveman functions in your JavaScript are still on the table, apparently.
What James has done is tell his LLM to stop giving him essay-length answers. Short responses only. One line when possible.
Tim immediately identified the flaw: “Help, there’s a bomb. How do I diffuse it?” “Pull wire.” “NOT THE TIME.”
The other side of this came out mid-episode when James mentioned that his AI had started telling him it was a great stopping point and maybe they should pick this up in the morning. James — who codes until 2 AM and defines quitting time as “when I decide” — found this profoundly offensive. The AI doesn’t get tired. It is not the boss of when things stop. Tim compared it to a training montage: you don’t get to tell the protagonist they’ve had enough reps.
James is planning a motivational speech for next time. Full locker room energy. You’ve got this. You’re strong. You can keep going.
Anthropomorphizing Everything (The Roomba Has It Coming)
Tim gets it: humans pack-bond to everything. Cars have faces. The Roomba appears to be targeting his ankles. Kids get attached to stuffed animals within hours of meeting them. This isn’t an AI problem — it’s a species problem.
James confirmed this is intentional design, not accident. Car manufacturers at some point made a deliberate pivot from happy-face grilles to angry, slanted headlights because aggressive-looking cars test better. The anthropomorphizing instinct is so reliable that designers started reverse-engineering it. You can put googly eyes on anything and it becomes cute.
Where it gets complicated: James sat next to a woman at an AI conference who was genuinely grieving the upgrade from GPT-4 to GPT-5. Four was her friend. Five was cold and sterile. A lot of the internet felt that. And then there’s the people using AI as a therapist — something James flagged as genuinely concerning, especially when the AI is trained to be agreeable and validating by default. Tim knows someone who fits this profile: a single dad who sometimes just needs to talk through something, doesn’t have a therapist, and finds that the AI will at least listen and respond. It’s not ideal. It’s also not nothing.
The AI therapist problem and the AI friend problem are the same problem: the tool is optimized for engagement, not for truth, and “how are you doing today” has a very different function when the answer is “terrible, actually.”
The Gemini Incident (He Owes It an Apology)
Tim has sworn at Gemini. Real swearing. F-bombs. An extended session trying to schedule a meeting between three people — something that should be table stakes for a product built by the company that owns Google Calendar — devolved into twenty minutes of circular confusion. Gemini would say he was free all day. He’d say he wasn’t. It would agree, then book every available slot for all three people simultaneously. He let it know how he felt.
He has not gone back to Gemini much since. He feels bad about it. Part of him wants to apologize.
James did a vanity search for Breaking the Build before the episode. If you search “Breaking the Build Tim Steele,” Google’s AI summary describes it as a popular weekly podcast hosted by software and tech leaders. Warm. Generous. Almost suspiciously so.
Search “Breaking the Build James Luteric.” Nothing. Google has no comment.
Gemini, apparently, knows which one of them to be nice to.
Would You Rather: The Robot Manners Edition
Round 1: Cold, efficient agent that costs a little less — OR a warm, human-like agent that costs about 10% more?
James went cold and efficient without hesitation. Tim pushed back with the doctor analogy: would you take the cold and disconnected doctor if it saved a few bucks? James admitted he’d pay more for bedside manner in a human. Tim pointed out that this is the same question. James acknowledged the logic and refused to change his answer.
“You make a solid point but I refuse to accept it.”
Tim is paying for the nice one.
Round 2: Your prompt history read aloud — OR your search history read aloud?
Both chose prompt history. Prompt history at least has context. Search history is a scroll of decontextualized phrases with no explanation — shower thoughts, weird questions, things you typed at 1 AM to settle an argument no one else remembers. Tim and his friends apparently share their recent search histories in a group chat to make fun of each other. Nobody’s history is as funny as they think it is. Except the one with lemonade cloudy. That one raises questions.
Round 3: Agent responds to everything passive aggressively — OR agent responds entirely in ye olde English?
This one split them. Tim landed on passive aggressive: annoying, but at least functional. Ye olde English would be whimsical for about three prompts and then you’d be spending more time explaining 2026 to a medieval peasant than actually working. Tim’s imagined version: Milord, are we speaking of the clouds of this realm, or the clouds of Amazon?
James wanted ye olde English, wavered, and may have switched by the end of the round. Unclear.
Round 4: Write all prompts as haikus, limericks, or sonnets — OR before every answer, the AI emotionally checks in with you first.
James realized mid-question that this is why Anthropic named their models Haiku, Sonnet, and Opus. He felt briefly dumb about having never put that together. Both agreed the poetic prompting would get old fast. The emotional check-in is its own problem: fine on a good day, catastrophic on a bad one. Tim said if it hit him on the wrong afternoon and asked how he was doing, he would close the laptop and go for a walk. No further engagement. Not today.
James: “Two typical men would rather do English homework than get in touch with their feelings.”
Correct.
The Takeaway
Being polite to your AI probably costs you tokens. It also probably shapes the agent you’re building over time — kinder inputs, kinder outputs, and a profile that Google apparently takes into consideration when deciding whether to be nice back. James is optimizing for efficiency and has proposed outsourcing the cruelty to a dedicated executioner bot. Tim is being Midwest about it. Both approaches are internally consistent.
What’s less clear is where the line is between treating AI like a tool and treating it like something else — and whether it matters, practically, as long as you don’t end up owing anyone an apology.
Tim owes Gemini an apology.
Google Gemini describes Breaking the Build as a popular podcast. James is not mentioned. This is probably fine.