Become the designer you want to be. Decide, build and prove with AI.

You leave with a working AI Skill, the assistant you built and evaluated, live at a link.

System prompt · Context window · Grounding · Retrieval · Memory · Tool calls · Permissions · Evals

1. 01How an AI system is put together

   Someone in planning asks whether the AI can remember what the user picked. You need to know if that is the model, the app, or neither. Get it wrong and you design the wrong settings screen, then the wrong error state. By the end you can answer that out loud.

2. 02Context and instructionGrounding, retrieval and memory

   More context is not better context. You rewrite one instruction twice, then add 2 documents it does not need. The answer gets worse while you watch. The assumption never comes back.

3. 03Tools, workflows, agents and permissionFailure, recovery and a first evaluation

   Your product will soon take actions without asking. Someone has to draw that line. You take 4 AI actions and decide which ones need a human first. Reversible is not the test on its own. You leave naming one action you would never let it take alone.

You leave with an evidence file, a friction score, and a persona nobody takes on faith. Plus a ranked plan carrying its rejected list.

Evidence file · Friction score · Personas · Opportunity map · Hypothesis backlog · Rejected list

1. 01Design AuditBrief · Competitive Analysis

   Any model can audit a product now. Almost nobody can make two raters agree. You score the product task by task, then write the protocol down. Another rater follows it and lands within 1 point of your total. That is what makes the number in your case study survive.

2. 02Persona CreatorAI Opportunity Map

   A model writes a persona in seconds, and it will sound right. You write the screener, run it, and rank drives and blocks against what people said. 2 to 3 cards, each block traced to the screen it happens on.

3. 03Hypothesis Backlog

   Ideas are cheap to build now, so the argument moves to which one. 3 audit findings become hypotheses, 7 fields each, ranked by effort against impact. Then someone else tries to disprove yours using only what you wrote. Most of the course hangs off this one file.

You leave with a design system with its foundations, one fully specified component, and a prototype that runs.

Primitives · Semantic tokens · Component tokens · Component states · Handoff spec · Prototype

1. 01Requirements & AcceptanceFlows & Structure

   The build happens overnight now, and nobody asked you anything. Your rules were the only steering input there was. Here you write the rule and the test that proves it, side by side. Every AI surface gets 4 defined behaviours: success, uncertain, empty, failure.

2. 02Wireframes & Copy

   The model draws all 5 states now: default, empty, loading, error, uncertain. It cannot decide what uncertain should mean in your product. You write that string twice, once badly and once useful. Only you know what your product can afford to be wrong about.

3. 03Visual Design, System & Prototype

   A flat frame cannot show a model being wrong in front of a user. Your prototype runs the 4 AI behaviours at real latency. Foundations first, then the real screens, then the parts harvested from them.

You leave with a test protocol, verdicts on your own hypotheses, and a decision you can defend.

Test protocol · Task success · AI evals · Impact measurement · Improvement loop

1. 01Usability Testing

   Watching one real person fail is the last research a model cannot run for you. Every task comes from a hypothesis you wrote on day one. Findings come back as 3 verdicts: confirmed, disconfirmed, inconclusive. You stop defending taste in front of people who outrank you.

2. 02AI Eval

   Anyone can make the screen now. Setting the pass mark is the rarer job. You build a case file from the rules you wrote in Build. The pass mark gets set first, because the same prompt answers differently each time. Known failures go into a regression set so they cannot return.

3. 03Measurement & ImpactImprovement Loop

   Nobody tracks whether users override the model. That number is about to matter. You name the metric before the design exists. Then you instrument the 4 signals: override, rephrase, abandon, ignore. Reading a metric you picked afterwards is decoration, and everyone can tell.