Prompting Is Fishing, Not Carpentry

By Aaron Adler

You can describe exactly what you want and get something close, or you can learn to read what the model is giving you, and let the best outputs lead.

Most people come to AI video generation with a carpenter's mindset. They write a detailed prompt, expect a precise output, and feel like something went wrong when they don't get it. The frustration is real, but the mental model is the problem.

You are not issuing instructions to a system that will execute them faithfully. You are casting into water you can't see the bottom of.

The Model Is Not a CNC Machine

A carpenter describes a cut and gets that cut. The relationship between input and output is deterministic. Change one variable, get one predictable change in the result.

Video models don't work that way. The relationship between a prompt and its output is probabilistic. Two identical prompts will produce two different results. A small change in wording can send an output somewhere completely unexpected.

This isn't a flaw. It's the nature of the system. Understanding that is the difference between productivity and frustration.

You're Casting a Net, Not Cutting to Spec

We rarely generate one output per prompt. We run variations — sometimes dozens — before we make any decisions about direction. We treat each generation as feedback: what is the model gravitating toward? What is the grain of this particular prompt? What details does it consistently interpret one way, and where does it roam?

That's the fishing part. You cast wide. You pay attention to what comes back. You don't evaluate each output against a fixed spec. You read the set of outputs as a signal about what this prompt, in this model, wants to become.

A carpenter who doesn't get the cut they asked for has a problem. A fisherman who pulls up something unexpected stillhas a catch.

What Good Fishing Looks Like

Our prompting process involves building families of related prompts anchored around a core concept but varying in style reference, camera language, texture, and mood.

We're looking for convergence: where multiple variations produce something in the same territory, that territory is probably where the model is strongest for this concept. We're also looking for outliers: the outputs that landed somewhere we didn't ask for but are genuinely interesting.

The outliers are often the best raw material in the batch.

Knowing When to Keep What You Caught

This is the editorial judgment the workflow depends on. The question we ask is not "does this match what I asked for?" It's "is this good, and can we build around it?"

You have to know the project well enough to recognize when something unexpected is actually a better answer to the brief than the brief was.

Budget for the Cast

Generating one output per concept and moving on is carpentry thinking. The real work is the wide casting, the batch review, the editorial culling, and the pivot toward what the model is actually good at producing.

Prompt like a fisherman. Cast wide, pay attention, and stay open to a better catch than the one you planned for.