There are nine finished portraits in the Heroes, Off Duty series. What’s not on that page is the real count of generations behind them — a lot closer to a couple hundred than to nine. That ratio isn’t a confession of inefficiency. It’s the actual shape of the work now, and almost nobody talks about it honestly.
The popular story of AI-assisted creation is: describe it, get it, ship it. The real version is: describe it, get it wrong nineteen times in ways that are individually hard to articulate, and slowly work out — by studying the failures — what you actually meant in the first place.
Generating was the bottleneck. Now editing is.
For most of history, the constraint on making things was making them. A painter with a clear idea still had to spend hours executing it. That’s inverted now. Execution is instant and nearly free; what’s scarce is knowing which of the instant results is right, and being able to say why.
This is a genuinely different skill, and it’s worth naming because it’s the one you now get paid for. It isn’t “prompting better.” It’s editing — the unglamorous act of looking at twenty plausible options and rejecting nineteen for reasons specific enough to act on.
The specific enemy is “generic correctness”
The failure mode you’ll hit most is not ugliness. It’s competence with no point of view. Ask for “a detective with a magnifying glass” and you’ll reliably get something correct, atmospheric, and completely forgettable — because it’s the statistical average of every detective image the model has seen.
| The obvious answer (what you get first) | The keeper (what you’re actually after) |
|---|---|
| A detective inspecting a clue | A detective revealing a hidden world inside an ordinary stone |
| A hero drawing the sword | A hero returning the sword, choosing not to fight |
| A wizard casting a spell | A wizard sitting still while the room quietly changes around him |
| “Correct,” average, instantly forgettable | Specific, deliberate, unmistakably one idea |
You don’t reach the right-hand column by asking the model to “try harder.” You reach it by getting bored of the average enough times that it runs out of average things to show you. The rejection is the creative act.
Why the ratio isn’t waste
Twenty-to-one sounds wasteful until you realize the nineteen rejects are how you find out what you mean. Each one you turn down for a specific reason — “the light has no logic,” “this is generic fantasy, not this particular idea” — is a rep that sharpens the target. Skip that and you’re regenerating on vibes, hoping the fifth try beats the first without ever learning why one did. (This is the practical, hands-on half of a bigger point I made in AI Made Creation Cheap. Taste Got Expensive.)
So the number to be proud of isn’t how few you generated. It’s how precisely you can explain the ones you killed.
An editing practice
- On your next AI-assisted piece, deliberately generate more than feels comfortable before choosing. You can't edit a range you haven't seen.
- Reject out loud. For each option you pass on, say the one specific reason. If the reason is "dunno, worse," you haven't found your standard yet.
- When something feels "correct but boring," name that explicitly — it's the generic-average trap, and it's the single most common thing to push past.
- Write the least obvious correct version of your idea before you generate anything. Aim at that, not at the first thing the prompt returns.
- Keep a "rejects" folder for a month. Reviewing what you consistently threw out is the clearest mirror of your actual taste you'll ever get.