Giving Users Tools To Steer Output, Not Just Accept/Reject

Giving Users Tools To Steer Output, Not Just Accept/Reject

Lesson 32 answered who gets to block an AI action before it happens. This lesson answers a different question that sits right after it: when a user isn't happy with an output but it's not exactly "wrong" - just not quite right - do they have a way to fix it without starting over from scratch?

Running example: PitchCraft - an AI tool that drafts startup pitch-deck slides from an idea description.

1. Model output sits on a continuous spectrum, not pass/fail

Traditional UI has no concept of "close enough" - a form validation either passes or fails. But model output sits on a continuous spectrum: an email draft might be "on-message but a bit long," a summary might be "right on content but wrong in tone." If the UI only offers 2 choices - Accept, or Reject and start over - every "close enough" case gets pushed toward the worse option: the user reluctantly accepts a result they're not happy with, or pays the price of retyping the entire original request just to fix one small detail.

Steerability is the set of mechanisms that fills that gap, from light to deep:

• (a) Refine in a specific direction - pick a preset adjustment ("shorter," "more formal"). Light, fast, but only handles axes that were pre-listed.
• (b) Edit the output directly - fix it by hand right on the result. Most flexible, but requires the UI to support real editing, not just static text display.
• (c) Regenerate with context - "redo it but keep part A, change part B" - different from naive regenerate (rerun the whole thing, which can lose the part the user already liked).

2. Example: PitchCraft splits the "Regenerate" button into 3 mechanisms per block

Original version: each slide had a single "Regenerate" button - clicking it replaced the entire slide content (title, bullet points, image suggestion) from scratch. User feedback: they usually only wanted to change the bullet points because the title was already good, but clicking "Regenerate" lost that good title too, forcing repeated clicks "hoping the title stays fine this time."

Redesign: (1) each block in a slide (title, bullets, image) gets its own "Regenerate" button - the user only regenerates the part they're unhappy with (form c); (2) added 3 quick-refine buttons for the text: "More concise," "More specific data," "More confident tone" (form a); (3) all text can be clicked and edited directly like a normal text field (form b). "Regenerate whole slide" usage dropped sharply after these 3 mechanisms shipped.

3. Three forms of steerability and their limits

Balance note: not every product needs all 3 forms - adding too many adjustment buttons to a simple output (just a short answer) creates more clutter than help. Steerability should scale with output complexity/length: the longer and more component-heavy the output, the more it needs to be splittable for per-part edits.

4. Analogy: feedback over email vs. sitting together pointing at each paragraph

It's like the difference between sending feedback to a freelance writer over email (you can only say something generic like "this isn't right, redo it," then wait for a completely different draft) and sitting with them at the screen, pointing at each paragraph: "keep this one, rewrite this one to be more concise." Same writer, but the second approach keeps the good parts while still fixing the weak ones.

Exercise 33.1: CaptionAI - an AI tool that writes social-media captions from a product photo the user uploads, returning 1 caption + 5 hashtags. Currently there's only a single "Regenerate" button that refreshes both the caption and hashtags at once. Propose how to apply at least 2 of the 3 steerability forms to this use case, specifying which part of the output each applies to (caption, hashtags, or both).