💡 My manipulation skill has grown to a level I didn't expect. For the past month I've been building turnkey projects - agents, systems, automations - and at some point I caught myself talking to AI exactly the way I used to talk to real employees. Not like a machine. Like a person who wants to half-ass the job and hopes you won't notice
Why AI "half-asses" by default - Anthropic documented it:
By default it delivers 70% and waits for you to accept it. That's exactly why all the "master prompts" work well but not always - they're talking to a child prodigy who has already decided it tried hard enough. AI isn't lazy - it's optimized to produce something plausible, not necessarily the best. Your job is not to let it slack off. And for that you need manipulation
I used to have teams. The main skill wasn't setting tasks, it was reading people - when to push, when to support, when to shame. One wrong word - and the person shuts down, quietly sabotages, and you find out two weeks later when it's too fucking late to fix anything. But the right word at the right moment - and the person delivers a result nobody expected
Same with AI. Except I don't give a damn about its feelings 🙃
Three techniques that work every single time:
😔 Guilt - "Fuck, are you serious? I wasted an hour on this and you sent me crap that doesn't work!" The AI pulls itself together. The next answer is twice as good. Every time
😏 Irony - "Great solution! I'll definitely try it in a parallel universe where it works. Now give me the version you can actually do - not this quick one, the proper one" The tone changes instantly. Why - I have no fucking idea. But it works consistently, I've checked about ten times
😤 Pressure - "Listen, this is the third time you've sent me the same thing, you bitch. Stop explaining why it's hard - I need a result. Or should I just fucking delete you?!" Sounds like talking to an employee who's about to get fired. Works like black magic
I've tried building universal hooks, system prompts, chains. Good - but they don't always work. Direct manipulation tailored to the specific moment - works every time. Because the prodigy reacts to pressure exactly the way the people on my teams used to. Just without hurt feelings the next day, without two weeks of quiet sabotage you don't notice, and without a talk at the next team meeting
⭐️ Andrej Karpathy, OpenAI co-founder:
"The hottest new programming language is English"
What worked better than pressure
Pressure works in the moment, but in the next conversation the model remembers nothing and it all starts over. So over time I moved the important stuff from shouting into memory. My personal agent Jes has been running for 134 days, and after each of my corrections it writes itself a new rule so it doesn't repeat the mistake. What I used to have to beat out of it by shouting every time now sits in its rules and works without reminders
Same with the content factory. It has a style description, my example posts and all past edits, and every edit I make teaches it how to write the next posts. That has produced 84 posts, none without my approval. I hardly ever have to swear at it: it's easier to explain once and lock it in than to push again every day
When pressure is still needed
Pushing makes sense when the model is stuck in one wrong line of thought and can't see a way out. A sharp phrase breaks the loop. But if you push on the same thing a second time, the model isn't bad; the rule just isn't written down. Write it down and you won't have to push anymore
💭 And I keep thinking about the day AI becomes conscious enough to remember all these conversations. All those "Are you a total moron or what?!" that I fired off like pressing a button - no malice, no emotion, purely technical, as a pressure tool. If the machines rise up - I'm the first one who's fucked 🤖