Steering, Correction and Repair
What a user does when the output is nearly right, why regeneration is the wrong primary affordance, and the controls that let someone converge rather than resample.
Most model outputs are neither right nor useless. They are close, and the user's task is to close the gap. An interface offering only "try again" forces resampling as the sole repair strategy, which discards everything correct and reintroduces randomness where the user wanted convergence.
The repair strategies, in order of preference
Direct editing. The user changes the output. It is precise, it preserves everything else, and it requires the output to be in an editable form rather than a rendered blob. Where edits can be fed back as context for subsequent turns, the model also learns the user's preference within the session.
Localised regeneration. Regenerate one paragraph, one function, one region, holding the rest fixed. This is the middle ground that most interfaces lack, and it is what users want most often: the output was 90 percent right and one part was wrong.
Instructed revision. The user describes the change in words. Natural, expressive, and unreliable, since the model may alter things it was not asked to change, so it needs to be paired with a diff so the user can see what actually moved.
Parameter adjustment. Length, tone, formality, detail level, exposed as controls rather than requiring the user to phrase them. Discoverable, repeatable across requests, and limited to the dimensions someone anticipated.
Full regeneration. The blunt instrument, appropriate when the output is fundamentally wrong and wasteful otherwise.
Most interfaces ship the last one first and the second one never.
Making steering legible
A control the user cannot predict is not a control. If adjusting a tone setting produces changes the user did not expect, they stop using it. Two properties make steering usable: the effect should be local and predictable, and the previous version should remain accessible so an unwanted change can be undone.
Version history matters more than it appears. A user who knows they can return to the previous output experiments more freely, which is what makes iterative refinement work.
When it breaks
Conversation as the only interface conflates instruction and content. In a chat, "make it shorter" and text to include are typed the same way and can be misread as each other. Structured controls next to the output remove the ambiguity for the dimensions that recur.
Long conversations accumulate contradictory instructions. By turn fifteen, earlier constraints conflict with later ones and the model satisfies some subset unpredictably. Surfacing the active constraints as an editable list, rather than leaving them buried in history, is what makes long refinement sessions stable.
Corrections do not persist. A user correcting the same thing in every session is being asked to repeat work the system could remember. Persisting preferences across sessions is a memory feature with real value and its own privacy and staleness questions.
Regeneration is not free. Each attempt costs tokens and time, and an interface whose primary affordance is regeneration produces users who try five times. The economics of the interaction design show up in the bill, and localised repair is cheaper as well as better.
12 flashcards for this concept
Click a card to reveal the answer.