🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.🚧 Under construction β€” I'm migrating this site from Framer to Next.js and publishing it early for testing, so a lot of the content is still in flux.
Open menu
Switch to Darkhello@product.inc
AIUX patterns

The character is a state machine

Tags: persona, identity, motion

An AI persona is not a logo. It is standing in for a process the user cannot see, which means the question is not what it should look like but which states it has to show. Design those first, then build one rig they are all paths through.

What the persona is actually for

In deterministic software, system state is readable off the artifact. The form is filling in. The file copy has a known total and a bar that moves against it. You rarely need a character, because the work is visible.

A product built on a model has neither. There is no known total to run a bar against, and frequently no partial artifact to read β€” the answer either exists or does not. So there is a gap between the request and the response where the interface has nothing truthful to show, and something has to occupy it.

That is the job. The persona is the readout for a process with no other visible tell. Which reframes the brief entirely: not "draw us a mascot", but "here are the states this process has, which of them does the user need to see, and what is the cheapest rig that can reach them".

Start from the states, not the face

Most processes have six. Most products need four.

Rendering diagram…

Idle and Working are the two nobody skips. Attending β€” it has your input, it has not started yet β€” is the one that removes the most anxiety for the least effort, because it separates "did that register" from "is it slow". Unsure is the one that is almost always missing, and it is the state that decides whether the persona is honest: a character that looks equally confident while guessing has been designed as decoration.

Cut the states the user does not need. But cut them deliberately, because a persona with one state is a mascot, and a mascot is a running cost with no job.

One rig, not a set of illustrations

The construction decision is what determines whether any of this survives contact with a roadmap.

A persona built as a set of illustrations β€” one drawing per state β€” is finished the first time someone adds a state and the illustrator has moved on. The states stop being added, the character drifts out of sync with the system it is reporting on, and within two quarters it is a static image that means nothing.

A persona built as one rig the states are paths through costs an animation to extend, not an asset. That is a difference between a change anyone on the team can make and a change that needs a budget, and it is the whole reason one of these is still telling the truth a year later.

It has a second payoff. If the rig is also the brand mark, then identity and status are the same object, and the product gets its most attentive branded moment for free β€” while the user is waiting.

Grounded in

DocRobot's mark is two brackets and two dots. A face, once you see it.

It blinks. It looks around. It squints while it works something out. It follows the cursor. Every one of those is a path through the same rig rather than a separate drawing, so the states cost animation time and nothing else β€” and the mark doing all of it is the logo, which means the thing carrying the brand and the thing reporting what the system is doing never had to be reconciled, because they were never two things.

The economy is the point, not the charm. A persona that needs a second asset per state is one that stops getting states.

Prior art, and where this differs

That survey is real and useful, and the products in it were chosen well. It is not a design question, though β€” documenting what other assistants look like cannot tell you when a persona is wrong, what states it owes the user, or how to build one that is still accurate after four more states have been added. Those are the parts that decide whether the character helps or just costs, and they are all downstream of treating it as a readout rather than a portrait.

When not to build one at all

Anti-patterns

The smallest version worth building

Write down the states your process actually has. Cross out the ones the user does not need to see. If two or fewer survive, stop β€” you need an indicator, not a character.

If four survive, draw the unsure state first and the working state second. Then find the simplest mark that can reach both, and check that a fifth state would cost you an animation rather than a commission.