Prompt steering is the idea that a prompt should be shaped while it is being constructed, not evaluated only after it is submitted. In most chat interfaces the model cannot steer you as you type (no channel exists for it to nudge, suggest, or signal understanding until you press Enter), and you cannot summon help mid-draft without leaving the prompt. Steering restores both directions, inline.
The gap it names
The Human Interaction essay makes the observation that humans never wait their turn to steer a conversation: nods, frowns, and backchannel sounds flow continuously while one person speaks, each sub-second, non-blocking, and on a different channel from the speaker's words. Current LLM chat is stricter than turn-based human speech: type, submit, wait, read. In LLM chat, the steering that makes human conversation efficient is missing exactly where feedback is cheapest to act on: while the user is still writing the prompt, the model has no channel to nudge the user, and the user has no channel to ask for help.
Why steer before submission
The Inline Prompting essay puts the economics plainly: a good prompt steers the model, and iterating on the prompt before submission saves the tokens that would otherwise be spent on back-and-forth with an expensive model. Tools to improve a prompt exist, but using one has meant leaving the prompt, running the draft through the tool, and copying the result back; the friction is high enough that most users skip the step and submit the raw prompt. Steering only works if it costs almost nothing.
Steering in both directions
| Direction | Mechanism | Example |
|---|---|---|
| System steers user | Inline cues: dimmed words with tips and alternatives | A vague word dims; navigating to it offers sharper alternatives |
| System steers user | Continuous grammar correction: single-concern rewrites, each revertable | Grammar fixes land dimmed as you pause |
| User steers system | Blanks: a _ summons assistance at a chosen position | improve prompt _ re-renders the draft in place |
The rule underneath
All of these obey the same constraint: the steering signal must not occupy the channel you are typing in. A dim, a note in the secondary display, and an in-place fill steer without blocking; a popup that takes focus interrupts. That constraint has its own name, the modality rule, and it is what separates steering from nagging. Measured on the 3-axis framework, inline steering costs one keystroke of start-up and zero end-lag, which is why it gets used where separate tools get skipped.
Related: The modality rule · Inline prompting · What is an inline cue? · GitHub repo