Pattern Library
Reusable AI UX patterns for steering, review, transparency, memory, exploration, and recovery.
Pattern Library
Reusable AI UX patterns organised by function. For each pattern: when to use it, when to avoid it, and the trust implications of getting it wrong.
Steering patterns — giving users control over AI behaviour
Parameter controls
Tone (Professional ↔ Casual), Length, Formality sliders
- Use: User needs repeatable, predictable AI variation
- Avoid: Controls are so granular they become the primary task
- Trust: High — explicit control signals respect for user agency
Style reference
"Match the tone of this listing description"
- Use: Qualitative requirements are hard to describe in words
- Avoid: Reference example is too far out-of-domain for the model
- Trust: Med — only works if AI uses the reference visibly
Scope constraints
"Only apply to the selected paragraph"
- Use: AI tends to go out of scope without explicit boundaries
- Avoid: Constraints are so tight they prevent useful output
- Trust: High — users feel safe delegating when scope is clear
Example-based
"More like this · Less like this" thumbs on outputs
- Use: Iterating quality requires many regenerations
- Avoid: Examples given are contradictory
- Trust: Med — requires accurate model interpretation of examples
Persona presets
"Research mode" vs "Quick answer" vs "Detailed report"
- Use: Product has distinct use cases with different optimal AI
- Avoid: Personas are vague and create false user expectations
- Trust: High if presets are honest about their differences
Negative prompts
"Exclude off-plan · Exclude above AED 5M"
- Use: User knows what to exclude more clearly than what to include
- Avoid: Over-constraining leads to low-quality bland outputs
- Trust: Med — verify exclusions are actually respected
Review patterns — meaningful human oversight of AI actions
Approve / Reject
Binary decision on AI-proposed changes. Simple, low-friction, binary.
High — makes control explicit. Best for: consequential single changes.
Diff comparison
Side-by-side or inline track-changes view showing exactly what AI modified.
Very high — changes are inspectable. Best for: document editing, data transforms.
Selective acceptance
User accepts some AI changes and rejects others at granular level (line, field, paragraph).
Highest — maximum control. Required when changes are varied in quality.
Staged rollout
AI changes applied to a small subset first (e.g. 10% of listings). Review before full rollout.
High at scale. Best for: bulk AI operations on production data.
Confidence-gated review
High-confidence changes auto-apply; uncertain ones surface for human review.
Med — requires well-calibrated confidence scoring. Audit regularly.
Time-delayed commit
AI changes queue for N minutes before applying, giving users a review window without blocking flow.
High — undo is always possible in the window. Best for: ambient AI.
Transparency patterns — making AI reasoning and action visible
Reasoning preview
"I ranked these by price growth because you searched this filter 4 times"
AI shows its logic before or alongside the output. For complex decisions, analysis, and research tasks.
Action log
Activity feed: 'Aria updated 12 listing descriptions · 14:30 today'
Persistent record of all AI actions: what ran, when, what it changed. Essential for agentic and ambient AI.
Source provenance
Source: DLD registry data · Last updated 6h ago
AI shows which data sources informed the output. For factual claims, market data, and property information.
Uncertainty signals
Yellow badge: 'Low confidence — please verify with an agent'
Visual treatment distinguishing confident from uncertain outputs. Always present, scaled to stakes.
What-changed view
Struck-through original text · New text in primary colour
Highlights exactly what AI modified vs the original. Track-changes pattern applied to any AI edit.
Decision trace
Collapsible step log with each agent decision and data source
For agentic flows: show the decision tree the agent traversed. What it considered, what it chose, why.
Memory patterns — AI that learns from context and history
Every memory pattern must be paired with an edit interface. Users must be able to read, correct, and delete what the AI remembers about them. Memory without inspectability is surveillance.
Persistent memory
AI retains facts across sessions.
Must be editable and deletable. Show the memory, not just its effects.
Session memory
AI tracks context within the current session only. Resets on close.
Lower risk. Communicate clearly when session context clears.
Preference learning
AI infers preferences from behaviour without explicit instruction.
Highest risk. Must be inspectable and correctable. Show inference source.
Editable memory
User reads, modifies, and deletes specific AI memories in a dedicated interface.
The most important memory safety mechanism. Non-negotiable for persistent AI.
Memory provenance
AI shows why it remembers something: 'Based on your last 5 searches for villas.'
Prevents the 'how did it know that?' uncanny valley feeling.
Contextual injection
Current workspace state fed to AI as live context, not stored memory.
Safe — ephemeral, no persistence, user controls the workspace.
Exploration patterns — supporting iteration and variation
Multiple variations
AI generates 3–5 alternatives simultaneously. User selects one to continue with. Prevents premature convergence.
When: Always. No AI output should have zero variation.
Branching history
User diverges from any prior output and explores that branch. Timeline of AI outputs, not just the latest.
When: Long creative or iterative workflows where direction changes.
Side-by-side
Layout comparing AI output vs original, or comparing two AI variations directly. No toggling required.
When: Diff review, quality judgement, copy/design selection.
Iterative refinement
Each output is a launchpad. AI accepts 'make this more X' instructions on any previous output in the history.
When: Quality emerges through iteration, not single-shot generation.
Progressive narrowing
Start broad (10 options), progressively filter to one. AI narrows with each round of user feedback.
When: Discovery tasks where user is still forming their intent.
Pinned reference
User pins a good output as a quality anchor. AI uses it as a target for subsequent generations.
When: Consistent quality is required across many AI outputs.
Recovery & control patterns — escaping and correcting AI actions
Undo AI
Every AI action is immediately undoable — keyboard shortcut, visible undo button, or gesture. No exceptions.
Impl: Treat AI changes like any edit. Cmd+Z must work. Always.
Checkpoint rollback
Revert to a specific named prior state, not just one step. 'Restore to before the AI rewrote the description.'
Impl: Store snapshots before every AI batch operation.
Manual override
User can always take direct control of any field or artifact that AI has modified, bypassing AI completely.
Impl: Direct edit affordances must exist alongside all AI controls.
Fallback mode
If AI fails or is unavailable, the equivalent non-AI flow must be accessible immediately without data loss.
Impl: AI is a layer, not a gate. Product works without it.
Correction interface
User can annotate what was wrong. AI records this and adjusts future outputs. Closes the feedback loop.
Impl: Correction is a first-class action, not an edge case.
Scope reduction
After unexpected AI behaviour, user narrows permissions easily. Granular permission controls, not binary on/off.
Impl: Permission control UI must be findable in under 10 seconds.