You can't mock a probabilistic interfaceYou can't mock a probabilistic interface

An interface with a model behind it has a range of outputs, and a frame in Figma shows one of them. The useful design question for an AI feature is what the range looks like at its edges: the slow response, the wrong one, the empty one, the refusal, the answer that is technically correct and useless. A design team at Notion AI has described a shared prototype playground on Claude Code where designers iterate in running software (Lenny's Newsletter episode), and Designer Fund's AI in Design 2026 report found that some teams now treat prototypes as their main output, with half of surveyed designers having shipped AI-generated code to production.

Five decisions that came from listening

A daily audio show on this site is written by a prompt, and the comments in the module that holds that prompt keep a dated record of design decisions, nearly all of which came from hearing real output.

On 2026-09-16 spoken hedging came out, including the reporting-verb qualifiers and the remarks about thin sourcing, because on air it came across as skeptical and confrontational even though the same hedges on a page read as careful editing. Filler words and performed emotion had been tried as a way to make the voices sound human, and they made the show sound more artificial, so they went. On 2026-09-22 the show went from two voices to five, so the voice itself carries the variety a back-and-forth used to carry. A hand-back and a one-line thanks went in at the end of every story so a listener hears that the story is over before the musical sting arrives. On 2026-10-27 the correspondent began opening each story by saying what it is about, because an opening that listed sources sent listeners back to replay the segment.

Every one of those is a fact about how output lands on a person at the moment of hearing it, and every one was invisible on the page, where the hedging read as good editorial practice, the source-first opening read as tidy, and the filler words read as realistic. A static transcript would have approved all of them, and the decisions surfaced only when the real prompt ran on real inputs and someone listened to what came out.

What a prototype for an AI feature has to hold

The real model call with the real prompt comes first, since in a feature like this the prompt is the design, and changing it should be as easy as changing a color. A saved set of ugly inputs comes second, because a designer's own test case is always a clean one, so keep a file that includes an empty input, a very long one, an ambiguous one, one in another language, and one asking for something the feature should refuse. Every state belongs on screen at once: loading, streaming, success, partial, failure and refusal, each a screen worth designing before the week of launch. The design system sits underneath, so the prototype is built from real components and stays close to production. The older advice on matching fidelity to the stage of a project still applies, and the fidelity question has moved into code, which Claude Code for design-to-code workflows covers from the designer's side.

A decision log in the prompt, and a test behind it

The prompt module records why each rule exists, with a date, so the next person or agent who edits it sees the history of what failed before. A test then holds the assembled prompt character for character to the text that aired, and a comment above it dated 2026-10-27 says the correspondent opens on the topic and the fixture changes only when the prompt is changed on purpose. A second test checks that every version of the shared listenability section tells the writer to say the point first and the reference after. The decision that came from listeners replaying a segment becomes a build failure if a later edit brings back the source-first opening.

Critique changes too

Reviewing a frame means judging a layout, and reviewing a prototype with a live model means judging a sample. Run the same input five times in front of the group and look at the spread: three runs in three different tones point at the prompt, the guardrails or the example set, and it takes thirty seconds to see. Read the worst of the five before anyone discusses the best, since the worst is the output somebody meets on a bad day.