//pragmatic leaders

Designing AI UX That Survives Reality

featured

A six-week path for designers now owning AI surfaces, where deterministic interaction instincts are no longer enough. It gives you a practical vocabulary for uncertainty, a pattern library for streaming/citations/confidence/recovery, and a way to design for hallucination as a permanent product condition. Built for teams who need to ship trustworthy AI UX without waiting to become ML experts.

forProduct designers, UX practitioners, and design leads responsible for AI features who can already wireframe and run research, but are being blindsided by non-deterministic behavior, trust failures, and cross-functional ambiguity.
outcomeYou can spec streaming, citations, confidence cues, regenerate, abstain, and undo behaviors; design around hallucination as a permanent property; run research that captures long-tail probabilistic failures; and produce a failure-mode storyboard plus UX spec engineers can implement.
6 weeks6 stages20 available items
ask if this path fits →

This path is for the designer who just inherited AI UX and realized conventional happy-path thinking is now a liability. If your team can demo well but keeps getting surprised in production, this is the reset.

Work the stages in order once. You will build a vocabulary for uncertainty, decide which trust patterns belong in your product, and learn research methods that surface long-tail failures before your users do.

The capstone is deliberately concrete: one feature, one failure-mode storyboard, one shippable spec. If engineering can implement your streaming, citation, confidence, abstain, regenerate, and undo choices without guessing, you did the work right.

01

Stage 1 — Designing for uncertainty is a new craft

Traditional UX assumes the system is deterministic and your edge cases are finite. AI UX starts with a harder truth — the long tail is the product, not a bug list. This stage resets your mental model before you touch a single component.

  1. 1When AI Is the Right Answer (and When It Isn't)manual13 min readStop adding AI to structured problems that needed a rule engine — this chapter prevents months of avoidable UX debt.
  2. 2Hallucination as a Product Problemmanual15 min readHallucination is not an outage to patch; it is a permanent interface condition you must design for from day one.
  3. 3UX Principles for PMsmanual9 min readRevisit first principles so your AI screens optimize for user confidence and recoverability, not just visual polish.
  4. 4Notion AI — Adding Intelligence Without Breaking Trustcase7 min readNotion earned trust with opt-in entry points and recovery paths — read it as a playbook for introducing AI without breaking core workflows.
02

Stage 2 — Build your AI interaction pattern library

Teams fail when every prompt screen is designed from scratch. This stage gives you reusable patterns and the judgment for when to apply each one — streaming, citations, confidence cues, regenerate, undo, and abstain.

  1. 1AI UX Patterns That Workmanual13 min readThe canonical pattern set for AI UX, with clear guidance on what earns trust and what only looks sophisticated in demos.
  2. 2Wireframing & Prototypingmanual9 min readTranslate probabilistic behavior into states and transitions your wireframes can actually communicate to engineering.
  3. 3Accessibility as a Product Decisionmanual10 min readConfidence cues and streaming outputs can quietly exclude users — this chapter keeps your AI UX usable under assistive tech and stress.
  4. 4Perplexity — Search Rewritten as Conversationcase6 min readCitation UX is not decoration; this case shows how trust rises or collapses based on source visibility and answer framing.
  5. 5Linear — AI as a Quiet Utility, Not a Chat Assistantcase14 min readLearn the opposite of flashy AI — quiet utility that saves time without asking users to trust a chatbot persona.
03

Stage 3 — Spec behavior, not vibes

A prompt is product behavior encoded in text, and designers need to shape that behavior explicitly. You are writing contracts for failure handling, safety boundaries, and response quality — not just copy.

  1. 1Prompt Design as Product Designmanual12 min readTreat prompt design like interaction design — every line should justify itself against a user job and a failure mode.
  2. 2Working with Designersmanual8 min readUse this to tighten collaboration rituals so PM, design, and engineering critique AI behavior with a shared language.
  3. 3Anthropic — Research Lab to Product Companycase7 min readThe lab-app boundary is a UX decision — this case shows how safety posture, refusal behavior, and product experience are inseparable.
04

Stage 4 — Run research on probabilistic systems

Researching deterministic flows misses the real risk in AI products. You need methods that capture confidence calibration, breakdown recovery, and what users do after a plausible but wrong answer.

  1. 1Customer Interviewsmanual12 min readUpgrade your interview guide from "did you like it" to "when did trust break, and what did you do next".
  2. 2Design Thinking for PMsmanual9 min readUse hypothesis-driven prototyping to test uncertainty handling early, before engineering hardens the wrong interaction model.
  3. 3GitHub Copilot — The First Real AI Product, and What Five Years Taught Uscase14 min readCopilot adoption was won over years of trust calibration — a lesson in measuring behavior change, not launch-day excitement.
05

Stage 5 — Design within real-world constraints

Your AI UX choices are bounded by model capability, policy, and legal risk in 2026. This stage helps you make defensible design calls when the right experience is constrained by safety and platform reality.

  1. 1Safety, Privacy, Compliance for Shipping Teamsmanual15 min readSafety controls are UX primitives — disclosure, redaction, escalation, and refusal patterns belong in your spec, not as post-launch patches.
  2. 2The 2026 Model Landscapemanual13 min readModel strengths and failure profiles should shape your interaction design — choose patterns that match what the model can reliably do today.
06

Capstone — Ship a failure-mode storyboard and AI UX spec

The path is complete when you produce one real artifact set your team can ship from: a failure-mode storyboard plus a UX spec for a live AI feature. Your spec must justify pattern choices for streaming, citations, confidence cues, abstain, regenerate, and undo using research evidence.

  1. 1Build an AI failure-mode storyboardcompassBuild the storyboard across best-case, likely-case, and failure-case runs so engineering can implement behavior, not mockups.optional
  2. 2Course 1: Model Selection and EvalscourseUse this optional drill to tie your UI decisions to model/eval constraints so your spec survives contact with production.optional
  3. 3Course 2: Prompt as SpeccourseUse this optional drill to sharpen prompt-level behavior contracts and make regenerate/abstain logic explicit in your UX spec.optional