UX essay · Agentic AI
Type / for a UX head start
Anyone at TrackBee can ask Claude for a screen now: developers, marketers, the people in support. What comes back works, but it rarely looks or sounds like TrackBee. So I built a team of four bots with agentic AI. Type a slash and they design, build, check and write with our UX rules, for everyone equally. They do not give a finished result. They give a starting point that still needs a lot of editing, and a person refines it from there.
Problem
The same components, turned down and then approved
Claude Code made building cheap. A developer can have a dashboard running before lunch, and a marketer can make a page without opening a ticket. What it did not make cheap is taste. Left alone, an AI builds the average of everything it has seen: grey cards, a sentence under every number, a badge for every fact.
Our Insights dashboard was built that way, and our product owner turned it down because it adds random text to explain numbers all over the place
. Two days later the same person approved a rebuild: Looks like an actual product now. Not vibe coded side project.
The components were the same, and so were the features. The approved one simply showed less.
-
Product ownerAdds random text to explain numbers all over the place.
-
Product ownerLooks like an actual product now. Not vibe coded side project.
That told me what was missing. Not components, since TrackBee already had a library. What was missing was the judgement about when to use them. Part of that judgement can be written down, and the rest stays with a person.
Team & audience
The roles moved, not the people
Cat Wu, who leads product for Claude Code at Anthropic, put the change in one line: on her team, designers ship code, engineers make product decisions, product managers build prototypes and evals
. Anthropic's own study of its staff heard engineers describe becoming leaders and managers of agents
.
The same thing is happening at TrackBee, with one difference. Most of the people who build with Claude here are not designers, and never will be. So the design has to come with the tool.
Deep diveFive kinds of work, and who does each one now
Boris Cherny, who created Claude Code, describes the work on his team as five kinds rather than five job titles, and any person can do any of them. Here is each one at TrackBee, for the designer first and then the product owner.
01Ideas and scope
Comes up with new ideas, many of them, knowing most will never ship.
At TrackBee this is the Designer: it draws the versions.
02Making it real
Turns a prototype into a real product, fast.
At TrackBee this is the Builder: it writes the code from the boards.
03Cleaning up
Cleans up the interface, simplifies the code and removes what nobody uses.
At TrackBee this is the Inspector: it flags what does not belong.
04After launch
Takes a product that works and keeps making it fit its users better.
At TrackBee this is the Inspector's questionnaire.
05Keeping it healthy
Keeps a grown system secure, reliable and fast.
At TrackBee this is the library and the four bots themselves, which I maintain.
Sources: Product management on the AI exponential, March 2026; How employment is changing at Anthropic, October 2026; Boris Cherny's five archetypes, posted on X in June 2026, paraphrased.
Process
Four bots, one loop
Each bot is a Claude skill: a long brief in plain sentences that Claude reads before it starts. Three of them hand the work to each other, and a person decides when the loop ends. The fourth, the Writer, is not in the loop: the other three read its rules. They also teach each other: what I correct at each step becomes one set of learnings, and it goes back to all four.
a page that shows which stores stopped tracking. No brief, no ticket template. They can ask the bots or ask me, and most ask me, because they know the quality will be higher.
- Lap 1Not ready: 3 blockers, 5 major, 2 minor
- Lap 2Not ready: 1 major, 2 minor
- Lap 3Ready, nothing blocking
An example run. The verdicts are written the way the Inspector writes them.
Deep diveHow the bots learn from you
The bots learn from you as well. What you adjust, prompt or answer at each step is a learning. The four bots come up with one set of learnings from it, and it goes back to each of them, so the next run starts from it.
What you teach at each step becomes one set of learnings, handed back to all four bots. Click a bot to see it.
What you teach at each step becomes one set of learnings, handed back to all four bots. Click a bot to see it. You adjust the design in Figma, and tell the Designer what to learn from it. You adjust the implementation by prompt, and tell the Builder what to learn from it. You answer the questionnaire, and tell the Inspector what to check next time. You correct the words, and tell the Writer which term or rule to learn. The four bots turn what you taught into one set of learnings. The set goes back to all four, so each starts its next run from it.The Designer draws in Figma, from the library only
The Designer works inside Honeycomb, the TrackBee library I look after. Every colour, corner and text style it places is tied to a library variable, so a change in the library reaches every board it has drawn. Before it draws anything it lists what the screen holds and gives each piece a verdict, and the default verdict is that it goes.
Then it draws each width TrackBee ships at, and writes a note above every board: what is hidden at that width, what a tooltip does, what a row opens. The Builder reads those notes. A decision that only lives in a conversation gets lost between the two. The words in its frames, and the figures it makes up, follow the Writer's rules, so the design and the build read the same.
Board note
Stores page at 1440. The menu is open. The table needs more room than this, so Status and Last order sent share one column until 1536.
At 1024 the table no longer fits, so stores show as cards and the cards and table switch is hidden. The menu stays open from 600 up.
At 440 the menu folds into a button, which carries the red count. Every card stays; they stack. A cut-off store name opens in a tooltip on hover, focus and tap.
The Builder turns the boards into code
The Builder reads the boards themselves, never a summary of them, and builds from Honeycomb's coded components. It applies the whole system in one pass, because the person asking rarely knows the system and will not ask twice.
It does not choose the words. That is the Writer's job, in step 04, and the Builder loads it before it writes a string. What stays with the Builder is where text may sit at all: a label, a value and a change, and no sentence under them.
The Inspector checks, and asks only what needs a person
The Inspector never changes code. It reads the build against the Builder's own rules, the Writer's rules and the Figma boards, and reports every difference with a file, a line and a fix: a component that is not Honeycomb, a colour that is not a token, a word that breaks the Writer's rules, a control the tablet board hides and the build does not.
What it cannot judge from code, it asks. It builds a questionnaire that puts each board next to the real, working build and asks a person to look, click and decide. On the Stores page that took 62 questions down to 14, because the Inspector answered 46 of them itself by measuring. On a phone or tablet it asks for whole tasks, not single screens, such as fixing a store that stopped tracking, in numbered steps that float beside the device while you walk them.
62 questions in the first version
14 left for a person, once the Inspector answered 46 itself
Deep diveAnswer the questionnaire yourself
When the person is done, the page writes the results as a prompt. Pasted into the Builder, it starts the next lap. This is a small working copy of that questionnaire. Answer it with the mouse, or with W and S to move and A and D to pass or fail.
Stores page · visual and interaction check
A replica, made for this page
Honeycomb Inspector
-
Look at the card next to its board.
Expected: the same look, spacing and weight as the design.
-
Hover the cut-off store name in the build, then Tab to it.
Expected: the full name shows in a tooltip both times.
iPhone SE · 667 × 375
-
Turn an iPhone SE on its side. The open menu leaves 323 × 285px for the page.
Expected: the stores can still be read and opened.
Checked by the inspector
- PassedOnly Honeycomb components
- PassedButtons in one row share a height
- FoundThe status colours its label instead of its icon
- PassedNo badge wraps to two lines in Spanish or Portuguese
- PassedEvery word and figure follows the Writer's rules
In the real loop, this text is pasted into the Builder and the fixes begin.
The Writer keeps the words in one place
The Writer is the fourth bot, and the only one that does not draw or build anything. It holds the rules for every word and figure that someone outside TrackBee reads: spelling, dashes and quotes, the names that never translate, how numbers, money, dates and time are written, and the agreed word for each piece of marketing vocabulary in Spanish, Portuguese and Dutch.
Those rules used to be written out again in the Builder, in the Designer's sample figures, in the Inspector's checks and in a style guide inside the app's repository. A fix made in one had to be remembered in the others. Now there is one copy. The other three bots load the Writer before they write or judge a word, and a rule fixed once reaches every screen after it.
Deep diveThe same values, written two ways
It also works on its own. It names the medium first, a screen, an email, a social post, a help article or a customer report, and then applies the rules that medium needs. So marketing, sales and support get the same words as the product. Below is the same handful of values twice: once the way an AI writes them when nobody tells it otherwise, and once the way the Writer does.
Numbers and dates are formatted live by your browser, for the language in the tab.
Hivemind puts them one slash away
Hivemind is TrackBee's shared Claude workspace: one repository holding the company's context and its skills, installed by everyone as a plugin. A bot added there reaches every colleague the next time they open Claude Code.
So nobody has to know the bots exist to use them. A slash lists them, and asking for a dashboard that looks like TrackBee is enough for Claude to load the Builder by itself, and asking for an email to a customer loads the Writer.
What shipped
162 changes in eight weeks, most of them corrections
The first version, a single skill for building, went into Hivemind on 18 September. It stayed small until the Insights dashboard was turned down. From then on every correction became a rule: the product owner's comments, a developer's review, a mistake I caught in a build. From the day the skill split into three bots, they changed 131 times in three days. The fourth bot, the Writer, arrived on 10 November, when the rules about words and numbers moved out of the other three and into one place.
- First version, one skill
- Insights turned down
- Rebuild approved
- Split into three bots
- Writer added
Conclusion
Some judgement can be written down, and some cannot
The bots also make TrackBee itself better. When a rule changes, every screen built after it follows, whoever asked for it. The Writer carries that into the words: fix a rule about dates or buttons once, and every bot and every email after it follows. It is the lesson of my earlier design systems taken one step further: a library is only alive when it is used as code, and now part of the judgement around it is code too.