Portfolio of Upamanyu (Upa) Das, a design leader with 14+ years of shipped product and teams behind 517M+ users. Head of Design at Octopus, building AI-native design orgs. Case studies on agentic design systems, AI workflows, and craft-led leadership.
Contact: upamanyu@upamanyu.in · LinkedIn
I’m Upa a design leader who is happiest with one foot in strategy and the other in the pixels. currently Heading Design at 🐙 Octopus, leading a ten-person design org spread across four time zones.
I have been designing digital products for the past 14+ years. The title and the work kept changing shape — but its core motivation stays the same: building others up and leaving both people & organizations better than I found them. My skill set is a blend of hands-on craftsmanship and building highly functional teams. Together, we create delightful products and drive business growth. I’ve started and grown 5 Product Design & User Research functions from the ground up, helped scale 3 mid-size companies, and managed Product Design teams of 25+ (ICs and leaders). I’ve delivered products across Consumer, Enterprise, and B2B environments, used by over 517 million users. My work has had direct influence on closing 21 record-setting deals, won back 15 lost contracts, and been responsible for 52% retention growth.
I trained at Sir JJ School of Art and the National Institute of Design, which between them beat two habits into me: draw it before you defend it, and question the brief before you answer it.
Over the last 14+ years I’ve built five design & research functions from the ground up — at BookMyShow, where we rebuilt one of the most complex booking flows in live entertainment; at Cleartrip, redesigning how millions book flights and hotels; and at TMNZ, re-imagining how New Zealand businesses pay tax. Along the way I was instrumental in closing 21 record-setting deals, winning back at-least 15 lost contracts, and growing retention by 52%.
I bring proven leadership, design expertise, and a proactive mentality to the table. I excel at facilitating cross-functional involvement and recognize that great products aren’t designed in isolation. My approach emphasizes building bridges between design, product, and engineering to create alignment that drives results. Most of my attention now goes to making design teams AI-native — the workflow, not the tool list.
I may not be a founder, but I am an entrepreneur. I invest in company success proactively, understand business fundamentals, and contribute to organizational vision. I treat every team I join as if it were my own company — measuring success in outcomes, not outputs.
I’ve spent the last decade maturing how design gets done — capacity planning, workflow systems, critique culture, and AI-augmented pipelines that compound team throughput. DesignOps is where strategy becomes leverage; I treat it as a primary leadership discipline, not back-office plumbing.
I write extensively, mentor younger designers, and volunteer in design communities. Giving back isn’t optional — it’s how the discipline grows. I’ve built and grown 5 Product Design & User Research functions from scratch throughout my career.
Titles never moved me off the tools and my core passion. I still take a design problem end-to-end most weeks — research, flows and wireframes — because a leader who stops making things slowly stops seeing them. The kit: Figma, ProtoPie, HTML/CSS, Claude with an MCP server or two, product analytics, and an unreasonable number of whiteboards.
Away from the screen I’m a wildlife photographer — twice named Natural History Photographer of the Year — and I’ve made two short films about New Zealand’s wild places: A Forest’s Dream and The Birds of Play. Waiting out a bird in the rain turns out to be excellent training for design leadership.
Rebuilding ratings and reviews on a mass ticketing platform: a rating construct that finally matched its own scale, a three-step form that took nine seconds, and a trust model that made the number worth quoting. Time on task 60s → 10s, reviews up 24%.
Re-architecting a mass ticketing platform from a transaction into a habit — a content tab, a discovery product and an ad server, funded by a template system that halved the design team and cut asset turnaround from a day to two hours.
Designing Group Optimiser for TMNZ: a three-step tool that nets a whole group's provisional tax position in one pass — transfers, purchases, sales, and the interest saved — replacing a spreadsheet-per-client ritual run against a hard IRD deadline.
A fintech design system rebuilt as instructions for a machine — 412 tokens in three layers, 26 components with specs an agent can read, and four maintenance jobs it now runs on its own.
Research is cheaper than redesign. I don't start pixels until we can say who it's for and what breaks if we're wrong.
Wildlife taught me this — you lose the eye the day you stop looking through the lens. I keep half my calendar for craft, whatever my title says.
Every design decision starts and ends with the people using the product. Deep user research and empathy aren't nice-to-haves — they're the foundation of everything we build.
An agent is fast, confident, and wrong in ways that don't announce themselves. I give it the repeatable work and hold on to the deciding — and I expect the same split from everyone on the team.
The best interfaces are invisible. I push teams to strip away complexity until what remains is intuitive, efficient, and delightful. If a user has to think about the UI, we haven't done our job.
The interface is the manual — nobody reads the other one.
If a flow needs a tooltip to survive, the flow is the bug. People don’t arrive wanting to learn your product; they arrive mid-task, distracted, and generous with nothing. The screen has one chance to say what this is, what just happened, and what happens next.
So I design for the person who reads nothing and notices everything. Every label earns its place, every state has an exit, and anything that needs explaining gets redesigned before it gets documented.
A leader’s core strenth lies in their ability to lead their team through failure.
It brings me joy to help people achieve their full creative potential.
My design leadership style is people first: people are foundational to building great things. With people at the center, I operate using what I’ve coined as the 4Ps of Leadership: Success for People, Partners, Projects, and Processes.
And this style is backed by substance: I’ve honed my leadership approach over many years of experience in multiple tech sectors, company sizes, and scenarios.
A film releasing on a Friday has about three weeks to make its money. Inside that window a rating is not a nice-to-have — it is the single most public signal about whether a stranger should spend an evening and a thousand rupees on it. Our ratings moved box office. Distributors watched them. And on some titles we held more of them than the two best-known international review sites put together.
We were collecting all of that through a form that asked for a hundred and forty characters before it would let anyone leave.
I ran design on this end to end — the research plan, the rating construct, the trade-offs I took to the branding and business teams, and the last pixel of the form. Twenty million tickets a month were being booked around it, so nothing shipped without a defensible reason.
Figma, Google Analytics, CleverTap, two rounds of survey research, a lot of phone calls, and Jira for the user stories.
The icon had escaped. At one point there were more than ten variations of the heart-and-star mark in production, depending on which surface you were standing on and which year it had been built.
The contribution went nowhere. Someone had borrowed the hundred-and-forty-character minimum from social media and made it a gate. So people padded — repeated words, unrelated text, a wall of emoji — just to get past it. And the people who genuinely wrote something thoughtful watched it disappear down the page with nothing acknowledging it existed.
Nowhere to say it. You could rate from the film’s detail page, or from a modal on the homepage. That was the entire surface area. Come out of a cinema, open the app, and there was nothing asking you the one question you were most qualified to answer at that exact moment.
And nobody had to have been there. Anyone could rate anything. Which meant a reader had no way of telling a person who sat through it from a person with an agenda — and there were plenty of the second kind.
The first finding came out of analytics, not interviews, and it was the one that reframed the project. The scale ran zero to a hundred percent in ten-point steps. The interface showed five hearts. Which means half the values on the scale needed half a heart — and the control gave you no way to express that, no affordance suggesting it was possible, and no feedback when you missed.
The distribution confirmed it in the least ambiguous way I have ever seen data confirm anything. Whole hearts spiked. The halves sat in the trough between them. People were not rating films more generously at eighty than at seventy — they were rating what the control could physically hit.
Then the qualitative half. One study with sixty-two users — survey plus phone calls, because the phone calls are where people say the true thing. A second with over fourteen hundred, drawn from our existing personas.
What makes people write? Sixty percent write when a film is either wonderful or terrible; eighteen percent write because they simply like writing; fourteen percent write to help someone else out. And why don’t they? It takes too long, and they don’t know what to say. Writer’s block, in a text field, about a film they finished forty minutes ago.
One line from a respondent did more to shape the final design than any workshop:
I’ve never written a movie review in my life — I’m not sure what to write, and even if I do, I don’t really know if anyone’s gonna get value from it.
Then we read three hundred-odd reviews properly. The discovery hiding in there was linguistic: Indian cinema-goers reach for the same short, loaded words over and over. Jhakaas. Blockbuster. Must watch. Paisa vasool. People weren’t short of an opinion. They were short of a sentence. We already had the vocabulary — thousands of people had written it for us — we had just never handed it back.
We pulled apart the end-to-end rating flows of the apps our personas already had on their home screens, looking for patterns worth stealing and dead ends worth avoiding. Nearly everyone lands on five points. One long-standing film database runs ten. Delivery apps show you the distribution as a graph. A ride-hailing app hands out compliments instead of asking for prose. And a social app had proved that a slider, of all things, is something people will drag for fun.
Branding had already settled the heart. Stars are the industry convention and I raised that as a usability concern, in writing, more than once — a heart carries an emotional read that a star does not, and people had learned stars for twenty years. That argument I lost, and losing it cleanly was the right call: the heart was by then a genuinely recognised mark in Indian entertainment, and throwing away that recognition to win a consistency point would have been vanity.
So the question became: what construct makes a ten-point scale honest, using a heart?
Existing — five hearts, half-heart increments, no affordance for the halves. Take one — ten hearts, which tests badly for exactly the reason Hick predicts: ten targets is a decision, five is a reflex. Take two — a plain range slider, playful and quick, but in testing people didn’t know it could be dragged. Final — a ten-point slider with a real handle, snap points you can feel, and haptic feedback on each step. It won every round, and it won on speed.
The form became three questions in progressive disclosure, in strict order of how much they cost the person answering. Rate it — one drag. Describe it in tags — optional, and pre-filled with their own vocabulary. Write something — optional, and by this point genuinely optional, because we already have a usable review without it.
Time on task for a tag-only review came in as low as nine seconds. And because it was built as one modular component, it dropped into every surface we could find: the detail page, the homepage nudge, a push notification three and a half hours after a showtime, the post-booking screen.
One trade-off I still think about. Eighty-eight percent of people folded the cinema into their rating of the film — the seat, the sound, the queue, the air conditioning. We designed a version that separated the two, and then cut it from launch, because a question you can’t act on is a question you shouldn’t ask. Splitting the score meant deciding what we’d do with a bad cinema, and who at that cinema we’d tell. That is a business loop, not a design detail, and it wasn’t built yet.
New touchpoints meant new rules about when it is acceptable to interrupt someone. We wrote them as user stories with our agile coach — given, when, then — which forced the awkward specifics into the open. Three and a half hours after showtime, not during. Within thirty days, not forever. Once.
Trust turned out to be four words. Booked on the app, sitting under a review, next to a rating. That is the whole intervention on the surface — and underneath it, the aggregate started weighting verified ticket-holders more heavily, so the headline percentage came to mean something closer to what a reader assumed it meant.
The fraud side got the unglamorous treatment: the backend learned the shapes that organised rating campaigns make — the burst, the identical phrasing, the account that only ever rates one distributor’s films — and started flagging them for a human instead of waiting for someone to notice on social media.
People don’t read, so the detail page stopped asking them to. Top five tags and top five reviews, in the position a shopper’s mental model already expects them — down the page, after the thing itself. The tags are the summary; the reviews are the evidence for anyone who wants it. In testing people skimmed the whole sentiment of a film in ten to twenty seconds.
One of the world’s largest marketplaces went on to list films with us, and our ratings were part of that handshake — which is a decent external audit of whether a number is trustworthy. Mentions and shares across social media went up alongside it, mostly because a tag is a shareable thing and a paragraph is not.
People don’t type. They will absolutely tell you what they think — they just won’t compose it. Give them their own words back and the contribution rate stops being a motivation problem and starts being an input problem, which is the kind design can actually solve.
People don’t read either. Every screen after this one I’ve designed on the assumption that someone will spend ten seconds and take away exactly one thing. Decide what that thing is before you open the file.
And check what the control can physically express. A scale with ten values and a control with five targets is not a rounding error, it is a data quality incident that ran for years and moved box office. Nobody found it in a review. We found it because somebody plotted the distribution and asked why it looked like a comb.
For twenty years the company had sold one thing: a seat. A seat at a small-town Christmas gig, a seat at an 80,000-capacity cricket final, a seat in row H for the Friday release. Fifty million people a month opened the app. Ten million of them bought something.
Everyone read that gap as a conversion problem, and every fix on the roadmap was pointed at closing it. I spent a month in the session logs before I’d say the other thing out loud: this wasn’t leakage to plug, it was an audience nobody had designed for — forty million people showing up every month for something the product didn’t sell.
I was Head of Design. With the Head of Product I re-architected the platform from a pure transaction engine into a content and engagement app — the pitch, the navigation model, the three new products, and the operating change underneath that made any of it affordable. Seventy people in the design org reported to me while we did it.
Sketch and Figma, InVision, Google Analytics, CleverTap, Jira, a CDN-side image pipeline, and an unreasonable number of printed storyboards.
Seeing the audience is free; getting a twenty-year-old business to build for it is not. Before a single screen existed I had to argue the gap was worth chasing at all. The revenue line was a booking fee on a ticket — works in our favour while cinemas are the default Friday night, but breaks down once streaming becomes the default every night. Three things made the case for me.
Streaming had quietly become the competition. Not another ticketing app. A phone, a sofa, and no seat to choose.
We were sitting on the country’s largest opinion set about films. More user ratings than the two best-known international review sites put together, on some titles. That number moved box office, and we were using it as a decoration on a listing page.
And the forty million. We called them non-bookers and dreamers, someone who wants the Friday show and can’t justify it this week. We had the stars, the comedians, the sportspeople and the production houses on speed dial. We had every lever needed to keep those people around.
Leadership gave us a quarter to prove the direction before it touched the roadmap. We watched what the forty million already did between bookings, then designed to it: three tabs, three moods. Home stays exactly where the muscle memory is — you came to book, book. It gains smarter filters and better discovery, and nothing else moves. Store is for when someone wants the merch, not the seat. Buzz is for the other twenty-nine days of the month, when nobody is buying anything and everybody is scrolling.
The restraint was the design work. Every instinct in the room wanted to make the home tab do the new job too. Leaving it alone is what let us ship the rest without a conversion scare.
Underneath, the business case is a loop rather than a line. Ticketing earns directly. Engagement doesn’t earn anything by itself — it earns by making an ad server worth building, and by putting a dreamer within one tap of a booking on a week they can afford it.
More content means more artwork. A lot more. Seventy people already sat in design: fifteen product designers, and fifty-five graphic designers, illustrators and print designers turning around key art, banners and posters for every show going live.
For a media company the obvious answer is to hire. Media companies always hire. I argued the opposite — that a design org this size was a symptom, and that the fifteen product designers were the ones who could fix it.
The hard part was never the tooling. It was moving a service culture — organisers ring their account manager, the account manager rings a designer, a designer opens Photoshop — onto a product culture, where we build the thing that does it and everyone uses the thing. I took that argument to marketing, to business, to ground ops and to the supply side one room at a time, using how the big international ticketing platforms run their creative pipelines as the proof it was normal.
Every creative request moved onto tickets. No more favours, no more corridor briefs, no more “just this once”. Painful for a quarter, transparent forever after.
The audit was ugly in the best way: eleven different image sizes in production for a single show, most of them the residue of a layout that no longer existed. We took it to two. One vertical, one horizontal, per show. Everything else — every crop, every density, every placement — got derived on the fly at the CDN rather than drawn by a person.
That one decision is what paid for the rest. Turnaround for the full creative set on a new show went from twenty-four hours to two. Not because anyone worked faster, but because most of the work stopped being work — it became a rule, running on request, on a machine that never got bored of resizing.
None of it was clever. It didn’t need to be clever, it needed to be consistent — which is still the part people get wrong when they hand work to something that isn’t a person. I wrote the guidelines as instructions to be executed, not read: exact ratios, exact safe areas, what to do when the title is longer than the poster, when to refuse and come back to a human.
An endless feed for the daily dose — celebrity noise, box office numbers, trailers, whatever happened last night. Highlights ride the top in a format everyone’s thumbs already knew. Underneath it, three or four content templates the editorial team could fill without a designer in the loop.
Time on the tab went from about ten seconds — the old sprinkling of content across the app — to roughly thirteen minutes.
The honest product. If streaming is where the country actually watches things, then be the place that answers what should I watch, and where is it. Tell us your languages, your genres and which services you already pay for; we’ll stop showing you things you can’t watch.
New product, so it started small — three or four thousand visitors — and doubled inside four months, peaking around twenty-four thousand. Average session, three to four minutes, which for a recommendation surface is a long time.
The part that had to be designed carefully or not at all. The rule I wrote and defended: a loyal booker never sees an ad. Ever. Someone who books every fortnight is the most valuable person in the building and the least appropriate person to sell attention against.
Infrequent users get personalised placements that push them back toward a booking. Dreamers — the forty million — get the broad inventory. Same card design as the rest of the app, so the units carry more context than a standard banner ever does, and the app doesn’t suddenly look rented out.
It was this platform that kept revenue moving through 2020, when nobody in the country could sit in a cinema.
Fifty designers became twenty. Nobody was let go — the rest were repurposed into ad-tech, Watch Guide and Buzz. Three products shipped in under two quarters on a codebase old enough to have opinions, with every change rolled out on a three-to-six-month adoption cycle because you don’t hard-cut an app fifty million people rely on.
Growth arguments are won on the supply side. The demand story — new tabs, new audience, new revenue — took one meeting. The eighteen months of actual work was convincing a service organisation to become a product organisation, and building the pipeline that made it survivable. Nobody puts that slide in the deck. It’s the whole job.
A big design team can be a design failure. Fifty-five people hand-making assets wasn’t capacity, it was an unsolved problem wearing a headcount costume. The fix wasn’t efficiency, it was deciding which parts of the craft were decisions and which parts were rules — then writing the rules down precisely enough that something other than a designer could follow them.
That distinction is the one I’ve spent every year since refining. The tools got dramatically better at the second half. The first half is still ours.
New Zealand runs provisional tax on fixed dates. Miss one, or guess low, and Inland Revenue charges use-of-money interest and late payment penalties from the day you should have paid. Tax pooling is the legislated way out: a pool of tax already deposited with IRD on those dates, from which you can buy a backdated credit that IRD treats as paid on time.
Most of the country’s accountants use it for one client at a time. The interesting problem is a group — a parent and its subsidiaries, filing together, some having overpaid and some having underpaid on exactly the same dates. The money to fix the shortfall is frequently already sitting two rows down the same spreadsheet.
Getting it there meant building a spreadsheet per entity, working out by hand who was over and who was under, and then raising each transfer, purchase and sale as a separate transaction. Against a hard deadline. In the busiest fortnight of an accountant’s year.
I led design at TMNZ, New Zealand’s original tax pooling provider. Group Optimiser was mine end to end: the discovery with tax agents, the flow, the interaction model for the netting table, the terminology, every state and every empty state, and the arguments with myself about how much of this to automate.
Figma with a component library and variables, ClickUp for the brief-in and real accountant sessions for validation.
Here is the ritual, and I want to be fair about it — it worked. It was just enormously expensive, entirely manual, and wrong in a way nobody could see until the deadline had gone.
Every entity in the group needs its position worked out: what IRD says is due on each date, what has actually been paid, what is sitting in the pool. Then someone reconciles those by hand into a group view, decides which surpluses cover which shortfalls, and raises the transactions one at a time. Add an entity and the work doesn’t grow by one — it grows by one against every other entity and every tax date.
The whole design came down to sequencing. An accountant does not want a settings screen; they want to be asked the handful of questions that change the maths, see the answer, and then commit it deliberately. So: preferences, then the optimised position, then confirm.
The rule I held to across all three: nothing is created until step three. Steps one and two are a quote you can run twenty times on a Tuesday and walk away from. That single promise is what let me put a live-recalculating table in front of people handling other people’s tax money without anyone getting nervous.
Every preference on this screen exists because leaving it implicit would produce a technically correct answer that the accountant would then have to unpick. Should tax that’s currently listed for sale be pulled back into the group? Should surplus that the optimisation doesn’t need be sold? Should tax dates that haven’t happened yet be included, when the client may still intend to deposit on them?
Each one carries a plain-English consequence and, where there’s a defensible default, the word (Recommended) — earned, not decorative. A recommendation on a screen about someone else’s tax liability is a promise, so each one had to survive a sentence explaining why.
This is the product. Every taxpayer in the group, their position at IRD, what they hold in the pool, and — the part that did not exist before — what the group can do for itself before anybody spends a dollar.
Include and exclude entities and the whole position recalculates in front of you: transfers out of the surplus entities, transfers into the short ones, the residual that genuinely has to be bought from the pool, and any excess that can be sold. Underneath it, the number the entire product exists to produce — the interest and penalties this group does not pay.
Two interaction decisions did most of the work. The table recalculates immediately, with no apply button, because an accountant tests a hypothesis by toggling — and an apply button turns eight hypotheses into eight round trips. And every derived figure carries its own explanation on hover, because “why is that number that number” is the only question that matters here, and a support call is a design failure.
Financial tools are judged on their edges, and this one had three that mattered. A taxpayer past the IRD deadline for pooling that year, who cannot legally be included no matter what the table would prefer. Tax dates in the future, which may or may not belong in the calculation depending on whether the client plans to deposit on them. And entities the accountant simply has not finished — not started, or half-filled — sitting in a group that is otherwise ready to go.
None of these are errors. All three were being handled as errors, in red, at the point of failure. I moved every one of them upstream into a stated exclusion with a reason and a way back, because an accountant discovering at step three that four entities silently dropped out is an accountant who never trusts the tool again.
The residual purchase still has to be paid for, and there are genuinely different products behind that: pay it down flexibly as cash allows, fix a fee up front and settle at a future date, or build an instalment arrangement. Three real choices, each with a different interest profile, presented as a comparison rather than a dropdown — because this is the one decision on the screen that is about the client’s cash flow rather than their tax position.
Then the AML confirmation, which is a legal requirement and reads like one. I fought to keep it in plain language and short, and lost about a third of that fight, which is roughly par for a compliance checkbox.
Group Optimiser doesn’t live alone. The dashboard has an Upcoming Deadlines tab where an agent tracks which clients are approaching which dates, and every entity carries a status: not started, incomplete, complete, pending approval, pending payment.
Running the optimiser is, by definition, doing the thing those statuses are tracking — so one of the step-one preferences quietly closes the loop and marks the optimised taxpayers complete. It’s the smallest feature in the project and the one agents mentioned most, because it removes the bit of admin that exists purely because two screens didn’t talk to each other.
Which is the version of automation I actually believe in: not a machine making the tax decision, but a machine refusing to make a human retype what it already knows.
Accountants do not finish at the screen. They finish when the client, the file and sometimes the bank all have the same document. So the same optimised position exports as a report designed to be read on paper: a group page, a page per taxpayer, footed with who generated it and when, and a payment page with the banking details on it.
In financial products the edge cases are the product. The happy path took a fortnight. Past-deadline taxpayers, future dates, half-finished entities, groups where the optimisation nets to exactly zero, groups with nothing to buy at all — that took months, and it is the entire difference between a tool an accountant trusts in June and one they open once.
Recalculation is a trust mechanic, not a performance one. The reason the table updates instantly isn’t speed. It’s that watching a number move in response to your own decision is how a person builds a model of what the system is doing — and a model is the thing that lets them sign off on it.
And say what you excluded, before you exclude it. Every quiet, sensible, correct omission a system makes on someone’s behalf is a small withdrawal from the same account. Spend it once and you have a support ticket. Spend it twice and they go back to the spreadsheet, where at least the mistakes are theirs.
Most AI-assisted design work runs the same play: a designer prompts a model, the model produces screens, tokens, copy. The AI makes things for you, once, and then the output sits there — a system that can’t look after itself.
We tried the inverse. Instead of asking an agent to design our product, I rebuilt the design system so an agent could work inside it — auditing tokens, catching drift, correcting components, keeping Figma and code telling the same story. The same jobs I’d hand a designer with a spare Thursday.
I led design at a fintech whose product does one unglamorous, enormously valuable thing: it helps small businesses pay their tax on better terms, usually through the accountant who already does their books. Money saved is the whole proposition — the platform has taken hundreds of millions in interest and penalties off small businesses’ bills.
Which makes the interface almost entirely numbers, dates, statuses and money. Those four are exactly where a design system either holds or quietly leaks. I owned the system, and I rebuilt it for a reader I hadn’t designed for before.
Figma, Storybook 10, Figma MCP, CSS custom properties, React + TypeScript, Claude as the agent.
A design system stopped being documentation for developers the day an agent could read it. If it’s now instructions for a machine, three things have to be true: the tokens must be semantically layered, every component must describe itself in a way a machine can parse, and the naming must carry enough meaning that an agent can make a judgment call without me in the room.
Worth testing on something real, not a demo with two buttons and a colour ramp. This system runs 26 components, dark mode, a state matrix that covers overdue payments and half-filed returns, and 412 tokens that all have to stay coherent while an agent works on them.
Three layers. Primitives hold raw values — hex, scales, font stacks — and no component is allowed to touch them. Semantic tokens alias primitives by intent: --color-text-primary, --color-status-critical. Component tokens scope to a single pattern — the savings figure, the payment schedule, the badge that says an accountant is acting on a client’s behalf.
The layering is what makes it machine-readable. An agent that meets --color-status-critical doesn’t need taste; it follows the chain — semantic to primitive to raw value — and lands somewhere defensible. Rename a primitive and nothing downstream notices. That indirection is the entire trick.
Here’s the chain the agent walks when it audits us. Flip the theme and watch the semantic column re-point while the component column sits perfectly still.
Dark mode is the cheapest honesty test a token architecture can take. If components only ever reference semantic tokens, dark mode is one attribute on the root element. If anything is holding a raw hex, it gets caught red-handed the moment the lights go down.
Storybook found ours. Two components had hardcoded near-black text — a statement label and an icon caption — invisible in dark mode and invisible in review, because nobody clicks every story. The agent found them by reading the stylesheet, not by looking at pixels.
The second structural bet, and the one that changed how my team worked. Figma MCP reads component descriptions — so those descriptions stopped being a sentence for a designer browsing the library and became a spec for something that has to rebuild the component from words alone.
“A card that shows a client’s tax position” is true and completely useless. What an agent needs is the surface token, the border token, the elevation, every prop and its accepted values, and what happens on hover. Write that down once and the agent stops guessing.
An agent will never leave a gap empty. Skip the disabled state and it will invent one — usually 50% opacity on everything, including the text you spent a week getting legible. Skip the error state on a payment row and it will reach for the nearest red it can find.
So the state matrix became non-negotiable: every component ships with every state drawn and bound to a token, including the awkward ones this product can’t avoid — a payment that’s overdue, a return that’s part-filed, a client an accountant hasn’t been authorised for yet. Completeness is what keeps the agent from improvising.
Two canonical files, deliberately split. One holds tokens — variables, text styles, the primitive and semantic collections. The other holds components and their machine-readable descriptions. The split isn’t tidiness: it means an agent asked about a colour queries one file, and an agent asked about a component queries the other, instead of trawling both and picking wrong.
With the structure in place I handed the agent four real maintenance jobs — the unloved kind that slip for a quarter because there’s always something shipping. Each one worked because the naming and the layering carried the context.
The part that matters: none of these followed a script. For the drift audit I asked one question — whether the stylesheet and the Figma variables still agreed. The agent decided on its own to enumerate the custom properties, pull the Figma collections, diff them, sort the gaps by whether anything consumed them, and hand back a list I could action. Structure gave it enough context to make those calls.
The big one: quality becomes measurable. Once every token has a semantic name and every component has a spec, “is our design system healthy” stops being a feeling someone defends in a review and becomes a number an agent reports on a Monday. The audit above wasn’t an illustration. It ran, and the system is cleaner for it.
The second: the craft moves upstream. The risk was never that agent-built screens look bland. It’s that an agent assembling from a sloppy system produces sloppy interfaces, faster, and with more confidence. The deliberate choices — a green reserved strictly for money saved, a red that only ever means a date has passed — are what make the assembled output feel designed. The agent is only as good as the vocabulary it assembles from.
And a quieter one, from years of leading design teams: briefing an agent well is the same skill as briefing a designer well. Clear intent, honest constraints, and the humility to let whoever’s executing make calls inside them. The teams taught me that long before the tools did.
The design system is no longer just documentation for developers. It’s instructions for a machine — and the designer’s job is to make those instructions worth following.