AI-native design practice · Alpha
Specify the judgement.
Then judge the output.
AI does not simply add generation to a design workflow. It creates new design work on both sides of it: making intent explicit before generation and reconciling output afterwards.
The important shift is not that the model can produce quickly. It is that the designer’s highest-value work moves from executing one answer to defining what a good answer is, then recognising it across many possible outputs.
A new step on both ends.
Teams often formalise generation while leaving judgement informal. That is where tone drift, hallucinated details, and superficially plausible work reach production.
Author intent before generation.
System prompts, persona documents, and escalation rules are design artefacts. They make the system’s judgement explicit before a model produces anything. The work is less like specifying a layout and more like briefing a very literal, very fast junior designer.
Reconcile output after generation.
Generated options still need to be checked against intent, accepted, edited, or discarded. Choosing the right result requires the same taste a senior designer once spent making a single option. Reconciliation must be a named, owned step—not an informal vibe check.
From execution to specification and judgement.
Production speed matters less as a differentiator when generation is abundant. Clear writing, strong taste, systems thinking, and comfort with ambiguity matter more.
Editorial taste at scale
Review many generated options quickly and still recognise which one is genuinely right.
Precision in specifying intent
Compress purpose, constraints, voice, and behaviour into instructions that stay useful across many users.
Ownership of ambiguity
Decide what “good” means before a complete specification or fixed interface exists.
Systems over screens
Design rules that can generate many valid conversations, states, and paths—not one predetermined flow.
Run the tools like a design system.
Treat prompts as design files. Save what works, record why, and iterate deliberately.
Use common things real users say. Review consistency across a set, not the quality of one demo.
Use generation tools for first passes, language models for language and logic, and review tools for accessibility, tone, and evaluation.
Design better conversation tools.
A conversational product is not good because the model is smart. It is good because someone made deliberate, visible decisions: a persona with constraints, rules for handing off instead of guessing, short testable outputs, and a metric such as correction rate that reveals whether the system is working.
The system prompt is a design document. The escalation rule is a design decision. The live log is the reconciliation step made visible.
That is the pattern worth carrying into workflow and hiring: do not let the new judgement work remain invisible simply because it does not look like a Figma file.
Return to the skills map