after 3+ years with these tools, the only thing that's actually compounded
i do this for a living, and i still can't keep up. nobody can. there's a new model every week, and a feed full of people insisting you're already too late.
so a lot of genuinely good creatives freeze. they bookmark threads they'll never read, hoard prompt packs they'll never open, and wait to feel ready, while people with half their taste ship every single day.
i want to make the case for stepping off the treadmill entirely. you can't win a race that has no finish line, so stop running it. build something that survives every lap instead, a practice.
what i mean by a practice
when i started, my process was embarrassing. type a prompt, hate it, change three words, run it again, forty times, then keep whatever came out least bad. i called that iterating. in reality, it was a slot machine.
GIF
This was the quality of my AI process in February of 2024 🤮
what really made things click for me was realizing the prompt is maybe ten percent of the job. the other ninety is a handful of repeatable techniques i run on anything, no matter which model is hot that month. here are three of them:
aim before you shoot
before i generate a single thing, i decide what the finished piece looks like and what my reference for it is. sounds obvious. almost nobody does it. once you aim first, every generation becomes a test of a decision you already made instead of a coin flip you're hoping pays off.
Everything I make now, especially in diffusion models like Midjourney, have layers and layers of planning and reference material.
learn to read the miss
when an output is wrong, there are two completely different kinds of wrong. a directional miss, where the whole vibe is off. and an instructional miss, where the vibe is right but the hands are melted or the text came out as gibberish. they need opposite fixes. most people panic and change everything at once, so they never learn which one they were even looking at.
For example, when making the cover for this article, I was running into some pretty tricky instructional misses - it wasn't making the logo as realistic as I wanted on the fish.
I ended up having to work backwards a bit and start on the fish independently first.
direct with references
the model follows what it can see far better than what it reads. ten perfect adjectives lose to one good reference image every time.
See how much better (and more accurate) we're able to get this image by using a reference photo instead of abstracting the prompt into JSON / loads of words.
why this beats chasing models
here's the part that matters. once you have the moves, every new release just becomes another thing to aim your process at, usually with improved prompt adherence, and that's the fun part. i take what already works and point it at whatever launched this week.
how to do it:
building this practice alone is slow. i spent years on it, and created a course to document by probably hundreds of hours of learning - it's called The AI Creative Director.
cohort 3 starts july 1 and closes the 6th, and enrollment's open now. the people who come through it are kind of ridiculous...folks making work for meta, google, microsoft, adobe and figma.
if you've been feeling behind, come build the thing that makes "behind" stop mattering.
