AI Text & Code Editing UX Research

How Leading AI Editors Show Proposed Changes

Inline diff with color-coding: When using Inline Edit (Cmd+K), Cursor shows additions in green and deletions in red before acceptance
Cursor Flow: Predicts next location + next edit, allowing users to Tab through a sequence of low-entropy changes (Tab pressed 11 times, other keys only 3)
Cursor Blame (Enterprise): Links each line of code to the conversation that produced it, distinguishing between Tab completions, agent runs, and human edits
Background editing: For larger edits, work happens asynchronously — users write pseudocode that gets “compiled” into full changes

Philosophy: “Should feel like an extension of your will. The moment you think of a change, minimal intent executes it instantly.”

Ghost text: Dimmed placeholder text at cursor for completions, accepted via Tab
Next Edit Suggestions (NES): Arrow in gutter indicates available edit suggestions for existing code (not just completions). Tab navigates to it, Tab again accepts
Inline diff for multi-file edits: Changes applied directly to disk with pending status tracked. Squared-dot icon in Explorer and tabs indicates pending edits
Accept/Discard widgets: Hover over changes to accept/reject individually or all at once
Working Set concept: Users explicitly define which files can be edited, giving precise control over scope

Cascade Bar: Dedicated UI showing which files have been changed by the AI agent
Continuous context awareness: Memories track codebase details, Rules maintain code patterns
Tab to Jump / Tab to Import: Suggestions go beyond code to include navigation and dependency management
Turbo Mode: Auto-executes terminal commands to keep developers in flow

Pattern	Description
Unified vs Split views	Let users toggle based on preference
Intraline highlighting	Show character-level changes within modified lines
Whitespace detection	Auto-collapse whitespace-only changes to reduce noise
Collapsible context	Hide unchanged code by default, expand incrementally
Gutter indicators	Color-coded margins (green/red/yellow) show change status at a glance
Navigation arrows	Up/down controls to jump between changes
Inline comments	Feedback directly on specific lines within the diff
Keyboard navigation	Full keyboard support for accessibility (arrow keys, Enter to accept)
Source control integration	Staging changes auto-accepts pending AI edits
Sensitive file protection	Approval prompts with diff preview for config files (`.env`, `.vscode/*.json`)

Key insight: Context management is critical — showing enough surrounding code to understand changes without overwhelming the reviewer.

“We’re still in a command-line moment, wearing wizard hats just to ask a machine for help.”

The article traces computing UX evolution:

AI editors are transitioning from #2 to #3 right now.

Principle	Implementation
State intent, not method	Users say “what they want,” AI figures out “how” (Manus approach)
Embedded, not external	AI lives inside the workspace, not in a separate chat window
On-demand initiation	Let users request suggestions rather than constantly interrupting
Micro-suggestions	Sentence/line-level changes, not full rewrites
Voice anchoring	Maintain the user’s style rather than homogenizing
Provenance	Show why suggestions were made to build trust

AI assistance can reduce users’ sense of authorship (–0.85-1.0 points on a 7-point scale). Mitigations:

iPreview: Instant previews across Finder/Spotlight, syntax highlighting for developers
Fileloupe: Fast browsing without imports/library management

Pattern	Example
Ghost text	GitHub Copilot inline suggestions
Inline diff overlay	VS Code Copilot Edits showing changes in-place
Gutter arrows	Copilot NES indicating available edits
Side-by-side panels	Copilot Edits in Secondary Side Bar while Testing view in Primary
Cascade Bar	Windsurf’s dedicated file change indicator
Plan visualization	Cursor’s Mermaid diagrams for agent plans

Integrated browsing: Extensions view via Activity Bar (⇧⌘X)
Discovery features: Popular extensions, ratings, download counts
Search/filter: By title, metadata, categories (predefined list)
Detail pages: Full description, README, changelog
One-click install: Install button → Manage gear
Version management: Right-click for specific versions
Security dialog: Trust confirmation for third-party publishers (since v1.97)
Sync settings: “Do Not Sync” option per extension

The best prompt is no prompt: Next Edit Suggestions and Cursor Flow show the frontier is anticipating user intent, not waiting for instructions
Inline > Modal: Changes shown in-place with accept/reject controls beat separate preview windows
Scope control matters: Working Set concept lets users define boundaries for AI changes
Speed over perfection: Users can iterate quickly with Undo/Redo rather than waiting for perfect first results
Progressive disclosure: Start simple (Tab to accept), reveal power features (partial accept, multi-file navigation) on demand
Non-technical users need goal-level interfaces: “Build a login page” not “create auth.tsx and import bcrypt”
Trust through transparency: Cursor Blame, provenance, and diff views build confidence in AI changes