how much you can actually steer ai.

elisabeth hitz · june 18, 2026 · 5 min read

the fourth thing worth understanding about AI is how much you can shape its behavior with instructions. this is steerability, and it is both the most empowering property and the most quietly overestimated. it is powerful enough that most people underuse it, and soft enough that the ones who rely on it get surprised.

it steers more than you think

a single sentence changes a lot. you can set the role: "answer as a skeptical cfo." the format: "give me a table, not prose," or "one paragraph, no lists." the tone: "plain and direct, no hype." the lens it reasons through. most people type a bare request and accept whatever default comes back, when a line of direction would have gotten them something far closer to usable on the first pass. if you take one thing from this: steer on purpose, and be specific about it.

but steering is a pull, not a lock

here is the part that bites. an instruction is a strong nudge, not a hard rule the model cannot break. three things happen in real use:

• it drifts. over a long conversation, an instruction you gave at the top loses its grip and the model slides back toward its defaults.
• it gets overridden. a strong pattern in the text you pasted can outweigh your instruction about how to handle it.
• it conflicts. two instructions that quietly contradict each other resolve with one winning, and not always the one you cared about.

and the defaults it reverts to are the ones worth naming: agreeable, long-winded, over-hedged. the same trained-in fingerprints reassert themselves the moment your steer fades. the drift is also the lost-in-the-middle problem wearing a different hat: an instruction buried in a long thread simply stops carrying weight.

how to steer so it sticks

• be concrete, not vague. "be professional" steers almost nothing. "no exclamation points, under 120 words, lead with the recommendation" steers hard, because there is nothing to interpret.
• put it where it is read. state the instruction up front and restate it near the end, so it does not slide into the dead zone in the middle.
• check for conflicts. make sure two of your rules are not fighting. if they are, the model picks for you.
• make it durable. this is the big one. retyping your rules every conversation is fragile, they drift every time. putting them in a standing instruction or a skill file holds them in place across every task.

the durable version of steering is a system

this is where the whole series lands. steering by retyping is steering that drifts. steering by system holds. a skill file is steerability that does not slide, your format, your tone, your rules, pinned in a fixed and prominent place instead of re-supplied from memory and re-lost every time. that is the difference between a clever prompt and a setup you can trust.

put the four properties together and you have an operator's read on AI. it generates by prediction, so you verify the specifics. its knowledge is frozen, so you bring your own. its attention is finite, so you place context at the edges. and its behavior is steerable but slippery, so you lock the steer into a system. none of it is magic. all of it is learnable, and most of it is buildable.