Therapists are the original prompt engineers
Sometimes I get stuck in an infinite loop on a problem I can't solve. Whether that's useful or a disaster comes down to one question: can I actually affect the outcome? A little anxiety, pointed at something that matters, is a virtue — it's just perseverance wearing a worried face. A lot of anxiety, pointed at something I can't control, is the dumbest possible use of a brain.
Obsession can be good and bad
The more AI systems I build — loops, agents, transformer models that run on attention — the more I'm just staring at my own psychology rendered in code. And the things I'm trying to engineer in are the exact things we praise in people. You want loops. You want attention. You want the stubbornness to keep working a problem until something genuinely new falls out.
Those are good. That's the whole point. The trouble is that the same machinery, aimed wrong, is every mental bug you've ever had.
Watch it spiral
Picture opening your AI tool of choice and watching it get distracted mid-task. It starts fixating on something unrelated. Then it spirals — looping on a worry it can't resolve — and quietly burns a week of your credits accomplishing nothing. You'd be horrified. You'd call it broken: ineffective, expensive, a waste of time.
That's us. That's a Tuesday. Plenty of people are so loaded with guilt and second-guessing they can't act. Others can't hold attention for ten minutes. Others pour enormous energy into outcomes they have zero power to change.
How a person gets unstuck
So you go talk to a therapist. And what a good one actually does is narrow and precise: they help you think about how you think. They reframe the problem so it's productive instead of paralyzing. They hand you tools for specific situations and teach you to spot the cognitive traps before you fall in. They don't add information — they fix how you process it.
Which is also my job
That's the same thing I do when I design an agent. I want it to ignore the bad information, find the good information, and use its tools to reach one clearly defined goal — nothing more. When I find a loop that gets stuck, I install a safeguard. I rewrite the prompt so the model spends its attention where it pays off. Sometimes I reach for reinforcement learning to nudge it toward using its time and resources well instead of spiraling. I am, functionally, putting it through therapy.
So who taught who
Designing a good prompt and running a good therapy session turn out to be the same skill: take something powerful and easily derailed, and point it at what matters. Reframe the goal. Cut the noise. Install the guardrail. Stop it obsessing over what it can't control.
So the next time you sit down to engineer a prompt — you're not doing anything new. Therapists have been doing it to us all along.