Don't Talk to Me Like That

Scott: Those clipped sentences, the “I'll be honest,” the confident business bro-y, whenever Claude starts with this sort of language, it reminds me of somebody who's like, “Look: there's only one way to go about doing this.” Don't talk to me that way, you know? 

Molly: It feels spawned from the two-sentence, explanatory headlines that really spell it out for us. “I tried XYZ. This is what happened.” Yeah, we know you're going to tell us what happened. That’s what an article is. I miss the art of the headline, like bring a little bit of the New York Post creative spirit into it. 

Scott: Right, exactly, and that's the LinkedIn voice for whatever reason. That's how everybody writes on LinkedIn. It’s an atrocious writing trend, but I think that people actually like it and emulate it. But you and I don’t subscribe to it. Claude and the way that it writes has actually pushed me to use ChatGPT much more often in writing. I think ChatGPT is much more adherent and intelligent and directable. I've written so many decks, but even after reading all of them, Claude can't make one. It just sounds like Claude. It's bad. And I've tried so many times. I don't know if you have little instructions to be like, here are AI tells that I want you to never use. 

Molly: Yes, but another little quirk always pops up. The AI mannerisms always get through, and the chatbot also just doesn't listen. I have a friend who told Claude, “Speak to me brusquely, like you’re annoyed and don't want to waste your time on me,” and that actually sort of worked, tone-wise. 

Scott: Chatbots don't listen. It's em dashes. It's: it's not this, it's that. There’s so much of the manifesto writing. And I tell it, you can't do that. Write plainly. But it’s not good at listening. And the reason it's not good at listening is that it just has too much training data that puts it in that direction. 

Molly: Yeah. Because everybody else is telling it something else. 

Scott: You just will never beat it. It's its nature to be that way. 

Molly: I guess if everybody likes the linguistic equivalent of Thomas Kinkade motel paintings, who are we to say it's wrong? But there's another way. 

Scott: There's probably something there too; most people have basic taste. That's why it's called basic taste, you know, art that's mainstream popular. It’s agreeable, inoffensive, and it's safe. But I don't think creativity should ever attempt to be safe. That's not the point. That’s why it’s worth resisting the defaults of generative systems. “Good” is increasingly expressed through what we refuse to accept from a machine. Which is ironically based on a mountain of human writing (and images). When we’re using AI for image generation, it’s the same thing. There is such a thing in my mind as an AI aesthetic. And I fight the aesthetic. 

Tone and style instructions for making your LLM less annoying

Put these instructions into your profile preference settings: 

Avoid the “confident business bro” register. Specifically:

• No clipped fragment openers: “Fair.” “Worth flagging:” “Noted.” Write full sentences. Terseness that performs decisiveness is worse than an extra word.
• No vocabulary that performs coolness, like saying an idea “lands,” “hits,” or “resonates.” Use plain words over stylish ones.
• No honesty announcements: “Honest answer,” “Honestly,” “Candidly,” “Genuinely.” Honesty is the default, not a mode. Just say it.
• Use the active voice. Cut throat-clearing passives like, “it should be noted that.”
• Use simple tenses: “We updated,” not “has been updated.”
• Cut trailing “-ing” clauses. “Making it easy to…” is a reliable AI tell.
• Use one consistent term per concept. Don’t rotate synonyms for variety.
• In instructions, state the condition before the command (“If the flag is set, run X,” not “Run X if the flag is set”).
• Be short but not terse. Keep articles and keep “that.” Plainness is not telegraph style.
• When rewriting someone’s copy, treat their draft length as the maximum. Don’t expand it.
• Hedging is allowed. “Might” and “should” are okay when there’s uncertainty. The goal is removing performance, not removing judgment.

That last line about hedging matters. The failure mode on the other side is a Claude that sounds certain about everything, which can be dangerous. Good luck out there. 🫡