Overview
How might we let people iterate on an answer without retyping the whole prompt or losing a good draft?
When to use
- Chat and writing assistants where users commonly ask for shorter, longer, or different tone on the same answer.
- Content generation where one-click transforms beat crafting follow-up prompts.
- Products like Gemini or Notion AI that expose modify-response actions on answers.
- Flows paired with regenerate, where refinement applies a directed edit to the current draft.
When to skip
- One-shot factual lookups where refinement chrome is unused clutter.
- Code patches where structured diff review is safer than vague “make it better.”
- When the product already supports natural-language edit-selection (magic edit) on the same surface.
States
Design the refinement loop on the answer itself, not only a hidden menu of generic improve buttons.
- 01
Answered
The first response is visible with contextual refine actions attached below or beside it.
- 02
Selecting
The user picks a directed transform: shorter, expand, formal tone, or try again.
- 03
Refining
The answer slot shows progress. Actions disable until the transform completes.
- 04
Updated
Refined text replaces the prior answer in the same surface. The original question stays put.
Key UX elements
The parts that must be present for in-place refinement to feel fast and safe.
Actions
Expose 3-4 named transforms on the answer.
Use specific labels like More concise or Add details. Avoid opaque Improve with no preview of intent.
Answer
Keep the full response in view.
Refinement updates the same answer block. Do not hide the prior text behind a loading shell alone.
Intent
Make each action describe the change.
Shorter, formal, expand, and retry should read as distinct transforms, not duplicate regenerate.
Progress
Show updating while the transform runs.
A brief in-place status prevents double taps and signals that the current draft is being rewritten.
Context
Preserve the original ask.
Refinement applies to the visible output, not a new prompt thread. The user question stays anchored above.
Rules
Refinement actions that replace history with no undo or version carousel.
Opaque “Improve” with no preview of what will change.
Refining silently changes cited claims without updating sources.
Too many micro-actions that bury Copy and Regenerate.
Evidence
| Product | Implementation |
|---|---|
| Gemini | Modify response actions for length, tone, and structure on answers. |
| ChatGPT | Regenerate and edit flows plus conversational refinement turns. |
| Claude | Rewrite and style adjustments via follow-ups and artifacts. |
| Notion AI | Inline improve/shorten/change-tone actions on generated text. |
