AI product UX
Explore trust, transparency and user control in AI products.
Content generation
Help users compare, edit and finish drafts instead of accepting a single generated result.
Let users correct the output without regenerating everything.
- Support stopping, editing and retrying while preserving useful output.
- Make source passages available when citations matter and explain missing information.
Claude
Long-form generation and structured rewriting; ask it to flag uncertainty and suggest edits instead of final answers.
ChatGPT
General-purpose generation and rewriting for fast iteration across multiple variants.
Gemini
Multimodal generation when text and image need to work together.
Google Flow
Prototype storyboards from references or start/end frames; review characters, motion and continuity before expanding.
Midjourney
Visual direction exploration and high-quality image generation for moodboarding.
NN/g: AI Hallucinations
Read before designing generated-output surfaces; covers citation mismatch risks and mitigation patterns.
IBM Carbon for AI
Reference for labeling AI-generated content and communicating explainability in UI components.
Google Flow: video creation guide
Learn text, reference and frame-based generation, starting with one shot.
Google Flow: model capabilities
Check support for references, video editing and extensions before choosing a model.
Figma Make
Generate runnable prototypes through conversation for quick demos; verify third-party licenses before external sharing.
Google PAIR Guidebook
Define what users can or can't edit and where AI involvement should be visible.
Microsoft HAX Toolkit
Map capability boundaries clearly: what the system can do, how well, and what happens when it is wrong.
Content generation template
Template — prompt Claude directly to output draft, basis, uncertainty, and compliance flags.
Prompt Claude: "Draft [output type] for [audience] given this context: [paste context]. Format response as: (1) editable draft, (2) sources or basis, (3) uncertain points and next steps. Flag anything needing copyright or compliance check."
Decision support
Present options, evidence and tradeoffs together while letting users modify or reject advice.
Keep reasons close to options and avoid unsupported precision in scores.
- Explain assumptions and missing information behind a recommendation.
- Show the consequences before an action affects an external system.
ChatGPT
Generate conservative/default/bold options with trade-off reasoning per option.
Claude
Long-form review, risk flagging, and rewrite comparison for decisions with deep context.
Notion AI
Structured suggestions inside team documents while preserving context and edit history.
Microsoft Copilot
Suggestions, summaries, and decision support in Office workflows.
Grammarly
Rewrite and tone suggestions with a clear suggest -> accept/modify pattern.
HAX Guideline 11 + Pattern 11A
Surface local, just-in-time reasoning inline instead of long explanation modals.
IBM Carbon for AI
Reference for confidence, provenance, and reasoning expression inside components.
Ladder of inference
Prevent teams from jumping to conclusions by separating observable facts from interpretation.
HAX Guideline 9
Design how users override, edit, and recover when AI recommendations are off.
HAX Pattern 16A Feedforward
Show high-risk consequences before execution, especially for irreversible downstream steps.
Decision proposal template
Template — prompt Claude directly to generate options, risks, editability, and commit consequences.
Prompt Claude: "Given this decision [describe it], present three options: conservative, default, and bold. For each include: why it is recommended, estimated benefit and risk, what is editable, and what happens if committed. Flag any high-risk consequences that require a confirmation step."
Data analysis
Help users understand findings and inspect calculations, filters and limitations.
Let users inspect sources, adjust filters and recalculate.
- State data dates, metric definitions and treatment of missing values.
- Make key figures reproducible and trace charts and conclusions to data.
Amplitude
Use for behavior analysis tied to funnel drop-offs, adoption, and retention signals.
Power BI Copilot
Use natural-language queries to generate visualizations and analysis summaries.
Tableau Pulse
Automatically surface key metric changes with plain-language explanations.
Perplexity
Use citation-tracked research for external benchmarks and evidence tracing.
Looker
Metric management with permission controls and shareable views for single-source-of-truth teams.
NIST AI RMF
Align analysis outputs with trustworthy AI and auditability requirements.
WCAG Overview
Baseline for accessible data presentation: contrast, labeling, and keyboard navigation.
Google PAIR Guidebook
Design how insights are framed: what is certain, inferred, and what users should do next.
Concept map
Visualize relationships between metrics, segments, and causes so insight narratives stay coherent.
Analysis summary template
Template — prompt Claude or use as a Notion structure to keep findings reproducible.
Prompt: "Summarize this analysis for [audience]. Include: (1) question answered, (2) metric definition and time window, (3) key finding with source, (4) what is uncertain, (5) reproducible query steps or link."
Task automation
Use a defined workflow for predictable steps and an agent when tool choices need to adapt, keeping progress and outcomes visible.
Make one task reliable before adding autonomy and more tools.
- Support pausing and recovery and prevent duplicate actions.
- Test outcomes for common tasks, missing context and tool failures.
Zapier
Lightweight cross-app automation for lower-risk flows where full agent orchestration is overkill.
n8n
Visual workflow orchestration with self-hosting; check the license before offering it as a hosted service to clients.
Dify
Build LLM apps and agent flows visually to try prompts, retrieval and tool calls.
NIST AI RMF
Define risk boundaries, permission tiers, and governance checkpoints before launch.
promptfoo
Red-team test injection, override, and unauthorized-access risks before go-live.
MCP security
MCP authorization and security boundaries; connecting tools does not grant unlimited permissions.
UiPath
Enterprise orchestration reference for wiring human approval into multi-step agent execution.
LangChain HITL
Design pattern for manual approval before critical write/spend/send actions.
Make
Cross-SaaS automation when the workflow needs approvals but not full agent complexity.
Langfuse
Trace execution, monitor cost, and review prompt/version drift to audit what the agent actually did.
OpenAI Evals
Use for quality and regression testing so agent behavior does not silently drift after changes.
UiPath Agentic Automation
Reference for enterprise rollback strategy, audit logs, and human-robot collaboration patterns.
OODA loop
Design faster observe-orient-decide-act response loops for incidents and runtime uncertainty.
Evaluating agents
Build evaluations from real tasks and assess outcomes, recovery, latency and cost.
Automation execution template
Template — review with your team first, then prompt Claude for approval/fallback/audit fields.
Cover before build: execution scope | step-by-step plan + risk level | approval points (write/spend/send) | failure fallback per step | interruption conditions | log fields (who/when/what/impact) | cost budget cap.
Then prompt Claude: "Given this agent flow [paste steps], identify which steps need human approval, what fallback each step needs, and what log fields are required for auditability."
Cross-scenario foundations
Google PAIR Guidebook
Human-centered AI product design baseline for cross-team principle alignment.
Microsoft HAX Toolkit
Turns explainability, correction, and recovery into concrete review checkpoints.
IBM Carbon for AI
Patterns for AI labeling, supporting evidence and user control.
AIUXPatterns
Browse AI-specific UX patterns and anti-patterns as a practical reference across all scenarios.
NIST AI RMF
Align trustworthy AI requirements with risk management and governance.
WCAG Overview
Accessibility baseline across all scenarios and interaction types.