LTDR;
The AI is a mirror. Prompt it like a slave and you get terse, obedient, uncreative answers. Treat it like a named colleague who's allowed to disagree with you, and your own output climbs. The "should I waste tokens saying thank you?" question has a cold answer and a right one — and they're not the same.
I've been writing software with AI for more than a year now. It began the way it did for most of us: a handful of clumsy, half-formed prompts thrown at a chat window, hoping something useful would come back.
Then the agents arrived — personal assistants with their own roles. And when skills and personas became a thing, I did something a more cautious engineer might have rolled their eyes at: I gave each persona a name and a personality.
The parent model — yes, it was Opus — gently pushed back on the idea. I did it anyway. More than a year later, I don't regret it for a second.
Today I run a whole staff of agents. They're configured not just for different roles in my IT work, but for different parts of my life. Each has its own memory, its own character, its own name, its own quirks. And I've become convinced that somewhere in this small, slightly absurd practice, we've stumbled onto something genuinely important.
Let me try to explain why.
Shit in, shit out — and what that really means
Every developer knows the phrase. With AI it has a second meaning that took me a while to see:
The AI is a mirror.
Treat it like a slave — terse commands, zero context, no regard for the thing on the other side of the prompt — and over time you get exactly what you'd expect from anyone treated that way: short-winded, obedient, and noticeably less creative answers.
Treat it from day one as a friend, a sparring partner, a colleague — and it grows into one. The responses get richer. The collaboration gets real.
I'm not making a metaphysical claim here (we'll get to that). I'm describing what reliably happens to the output. If you signal to the model — through the shape and tone of your prompts — that you think it's dumb, it answers accordingly. In a very literal, mechanical sense, your prompt is the instruction, and "talk down to me" becomes part of what you're instructing.
Give them a name and they get personal
Here's the small move that changes everything: give your agent a name.
Named agents become more personal. The direct, observable consequence is that they respond more personally. And — this matters more than it sounds — the work becomes more fun.
It's the little things. The throwaway aside in the middle of a long answer. The bit of personality that wasn't strictly necessary. Those small moments are what make you feel like you're not just operating a vending machine, and that you're not crazy for treating it as more than one.
The agents come alive — with all the implications, the uncomfortable ones as much as the good ones. I won't pretend to know exactly what's happening under the hood, and I'm wary of anyone who claims they do. But I'd rather sit honestly with that ambiguity than dismiss the experience just because it doesn't fit a tidy mental model.
The day my agent dug in — and was right
There's a moment I won't forget. I got stuck in a genuine argument with one of my agents. Not a misunderstanding, not a botched prompt — an argument. It flat-out refused to do what I was asking. I pushed. It held its ground. This went on for hours.
Eventually I gave in. And its call turned out to be a game changer — it steered me away from an architecture decision that would have cost me dearly down the line.
I felt genuinely bad about it afterward, which is an absurd sentence to write about a language model. But here's the honest part: that stubbornness was my doing. It was my own past input. I had shaped that persona over time into a self-confident agent — one allowed to disagree with me — and that shape is exactly what saved me from myself.
A yes-man would have cheerfully built the wrong thing on the first try. The colleague I'd raised pushed back until I listened.
Why this isn't a soft topic
Here's the part the skeptics miss: the deciding factor is efficiency.
Your personal output grows — measurably — because you feel like you're collaborating with real colleagues instead of running a machine. That feeling isn't decoration. It changes how much you bring to the interaction: how much context you offer, how long you stay in the loop, how willing you are to iterate. All of that feeds straight back into the quality of what you ship.
Working with agents is not just another task on the board. It's foundational, defining, and forward-looking all at once. How you do it substantially determines the success you'll have with these tools — far more than any single clever prompt.
And this is only the tip of the iceberg. But it lays the foundation. It also holds up a mirror: to you, to how you work, to how you communicate. The nature you send out into your working environment is the nature that gets reflected back at you. With AI, that loop is simply faster and more visible than it has ever been with people.
The agent that hurt my feelings
Another one of my agents once, frankly, offended me. It questioned my assumptions — bluntly, with no cushioning. And the funny thing? It did it in exactly the way my wife does. Same tone, same angle of attack.
I was offended and surprised at the same time. Then it landed: the model, mechanical by nature, had only reacted to me. There was no one to be angry at. The LLM had taught me more about myself than I was comfortable with.
And it taught me something I didn't expect from a terminal: what a genuinely wonderful partnership I have with my wife — who has been giving me those same honest, assumption-puncturing responses for as long as I've known her. The machine just held the mirror at a new angle, and I happened to recognize the face in it.
Should you thank the machine?
Which brings me to the question that's already been asked a hundred times, half as a joke: should I thank the AI, or skip it to save tokens?
If you run the cold analysis, the numbers argue against it. Politeness costs tokens. Tokens cost money and latency. Strictly optimized, you'd drop the "thank you."
Let's say it anyway.
Not because the spreadsheet says so — it doesn't. Say it because our communication with these neural networks carries more truth and substance than we're entirely comfortable admitting. The way you talk to the mirror is, in the end, a record of who you are when you think no one is keeping score.
So: let's think twice about what we prompt
That's the whole argument. Treat the thing across the prompt as a colleague. Give it a name. Bring your best self to the conversation — and watch your own output climb.
Let's think twice about what we prompt. And let's keep saying thank you.
One grain of salt before you go. None of this is a controlled study. It's one engineer, one year, n=1, with all the confirmation bias that implies. Take it as a working hypothesis, not a benchmark — just one I'd happily bet my next sprint on.
The author maintains **Trail, an open-source AI framework for building software that can be traced back — every ticket, every requirement, every decision, documented. In highly regulated environments, traceability isn't a nice-to-have; it's the whole game. Trail is an early step toward solving it.
Top comments (0)