Programming is design first, then execution. AI work is different — first brushstroke, response, paint over. A painterly improvisation. Today we walk slowly through this shift in perspective.
People who have used AI tools for a month or two all say the same thing. "I hear you need to write perfect prompts, but mine are always rough and the results are mediocre." Every time I hear this, I answer carefully. You're imagining AI wrong. Let me unpack slowly. Today is about flipping your relationship with AI at the root.
This essay covers three things: why the "perfect prompt" fantasy is wrong, what AI work actually is, and how to shift how you work.
Most people imagine AI as a "program." So they handle it like programming. If you give perfect instructions, you get perfect output. That's why "prompt engineering" became trendy.
The essence of programming is: draw the blueprint perfectly, then execute it. Like architecture. Column here, wall there, roof like this. Precise design, precise result. To change it, redraw the blueprint.
But AI doesn't work that way. AI gives a slightly different result every time. Run the same prompt 10 times. Ten different results. It's designed that way. It's not a program. It's a stochastic generator. That's not a bug. It's a feature.
So if you try to write perfect prompts, you're playing the wrong game. Working from the wrong metaphor.
To understand AI correctly, swap programming for painting. Picture a painter in front of a canvas.
The painter doesn't draw a blueprint. They make the first brushstroke. The color spreads differently than intended. Looking at it, they think "ah, this direction is better" and make the next stroke. Another unexpected result. They adjust again. Stroke and reaction talking back and forth until the painting comes together.
AI work is exactly this.
That's the essence. Not perfect instructions — flexible conversation. Not control — interaction. Not design — improvisation.
Don't program AI. Paint with it.
One-line summary.
The biggest gift of this reframe: when the result diverges from your intent, it's not a failure. It's an opportunity.
I hit this often when writing video scripts. I give direction A and AI returns direction B. Before, I'd say "I miscommunicated" and retry. Now I read the off-track answer first.
Because sometimes that off-track answer reveals a direction I hadn't considered — often a better one. Honestly, a big chunk of my popular videos weren't in my original direction. AI drifted, I caught the drift, and developed it. Of my 7 videos with over a million views on YouTube, 4 started this way.
When a painter's brush slips and creates an unexpected line, the great ones keep the line. That becomes the painting. AI work is the same.
To dance this improvisation well, three skills from three years of practice.
One — write the first prompt like a seed. Don't try to be perfect. Sketch the direction and throw it. "An essay about AI tools, around 5,000 characters, beginner audience" is enough. See the first response and start the conversation there. 3-minute prompt + 27-minute conversation beats 30 minutes of prompt-polishing, every time.
Two — don't evaluate the response, observe it. Many people judge "good/bad" instantly. Don't. First observe: "What part was surprising?" "What phrasing was new?" Observation leads to the next direction.
Three — change direction often. Ten rounds in one direction is worse than 3 rounds each in three different directions. The courage to say "drop this direction, try something else entirely" when stuck — that's the core of improvisation.
This improvisation sense needs practice. You can't suddenly declare "now I improvise!" and do it. I recommend a weekly AI free session.
Block one hour a week. No goal, no deliverable. Just converse with AI. Light topic — "let's talk about the book I just read." And hand the flow of the conversation entirely to AI. If AI drifts somewhere strange, follow. Then throw strange questions back.
Do this for 6 weeks and the sense forms. I tried it; I had 12 students try it; all felt the change. Because the practice teaches "how to let go of control." Without that, no improvisation.
One caveat. Improvisation is not license. Completely dropping the reins and saying "AI, do whatever" is not improvisation. You keep holding the direction, but stay open — that's improvisation.
A painter doesn't outsource brushstrokes like AI. Each stroke is their decision. They just observe the result with an open mind. AI work is the same. You are the subject of each prompt. You respond flexibly to the answer, but the right to decide direction stays with you.
When that boundary breaks, AI starts dragging you. When it holds, you dance with AI. Huge difference.
One downside of improvisation, to be fair. Unpredictable results mean good moments are easy to lose. Painters doing improv take photos often, recording each stage. So they can look back and see which point was best.
Same with AI. When a good response lands mid-conversation, I immediately save it to a separate file. Try to find it later and the context has shifted; you can't reproduce it. That's improvisation's nature. This small save habit is the hidden secret of improv work. Of my 7 successful videos, 5 owe it partly to this habit — not losing the good moment.
One honest note. Improvisation doesn't fit everything. There are two zones.
Improvisation zone: writing, planning, ideation, design concepts, video scripts, branding. Anything requiring creative judgment. Work with more than one right answer.
Wrong zone for improvisation: numerical calculations, data organization, translation accuracy, legal/medical facts, functional correctness of code. Work with one right answer. Improvise here and you break things.
For me, roughly 70% of tasks are improvisation, 30% are right-answer. Mixing them up kills results. "Perfect instruction" on improv zone = flat results. Improvisation on right-answer zone = wrong results. When a task arrives, decide which zone it's in first.
AI is not programming, it's painting. Not a perfect blueprint, but improvisation that starts with a first stroke. Results that diverge aren't failures, they're opportunities. Seed-like prompts, an observing stance, courage to change direction — these three are the dance.
Tonight, try one thing. Throw AI a much shorter, much rougher prompt than usual. Don't judge the first response — only observe. Start the next conversation from there. Small shift, but the output starts to change.
Don't program AI. Paint with it.
Remember — seed, observe, pivot.