Select → refine → transform
Alongside two other designers, I explored how designers and writers refine AI‑generated work. As part of Carnegie Mellon University’s UI for AI initiative, we designed interaction patterns that give people more control over refining AI-generated ideas instead of repeatedly starting over.
AI interfaces prioritize regeneration over controlled refinement.
Reframed refinement from a regeneration problem into a control problem.
Designed interaction patterns for targeted edits and version comparison.
Problem
AI tools generate quickly, but refinement often breaks down after the first draft. Users are left choosing between regenerating everything or editing manually.
AI generates a first draft
Needs refinement
↻ Generate again
Entire response changes
SelectEditRepeat
Slow and repetitive
How might we support refinement across multiple levels while preserving structure and user intent?
Research
I analyzed existing AI tools and refinement workflows to understand how users maintain control while iterating on generated content. Existing systems tended to prioritize either broad exploration or precise edits.
Users could refine one area without changing the rest.
Across our review of existing AI tools and user testing, three patterns consistently emerged. These findings became the principles that guided our interaction design.
Support multi-level refinement
Users naturally switched between broad revisions and precise edits.
Preserve structure during refinement
Participants wanted to improve selected content without losing surrounding work.
Keep users in control
People needed clearer visibility into what AI changed before accepting edits.
“Show me what changed so I can make the call myself.”
Instead of optimizing for better first outputs, we explored how users could refine AI-generated content while preserving structure, intent, and momentum.
Regenerate everything
Lost structure
Targeted refinement
Preserved context
My concept was selected because it supported both broad and precise refinement without interrupting workflow.
Broader changes like rewrite, merge, tone, or structure.
Highlight exactly what needs refinement.
Refine selected content without regenerating everything.
Key decisions
Three design decisions made AI refinement more controllable.
Users shifted between revising entire responses and fine-tuning individual words. We matched the interaction to the scope of the edit, so people could refine only the part they needed.
“It rewrote my intro without asking.”
BeforeWhole output regeneration
Regenerating full outputs often overwrote parts users wanted to keep.
AfterSelected-region refinement
Enabled targeted editing so users could refine selected regions without changing the rest.
Challenge
Users needed to compare versions and understand what changed.
Decision
Added side-by-side version comparison and navigation.
Impact
Users could judge AI changes before committing to them.
Solution
We designed a refinement workflow that combined broad edits, targeted changes, and version comparison so users could steer AI outputs without losing context or starting over.
Broad changes without rewriting prompts.
Refine specific regions without changing everything else.
Review iterations before committing to changes.
Impact
Maze testing validated that users understood the refinement model and successfully completed tasks across overall refinement, inline refinement, and comparison workflows.
100%
Task completion across core refinement tasks
Users completed overall refinement, inline editing, and comparison tasks successfully.
0%
Misclick rate during refinement tasks
No navigation or shortcut confusion appeared during testing.