The picture comes first. That is how I work, and it is why this started.
I was two days into auditing a single flow when I realized the audit was going to be wrong. Not in its findings. In its confidence. I could tell you what was broken inside that flow. I could not tell you whether we had already solved it three verticals over, or whether the fix I was about to recommend would collide with something shipping next sprint in a pod I do not sit in.
That is a leadership problem before it is a design problem.
The cost of shipping faster than you can see
Credit Karma ships continuously. Tests roll out, variants win, screens change. The product that exists on Monday is not quite the product that exists on Friday. Multiply that across five verticals and the math turns against you. A designer in Insurance can go an entire quarter without seeing what Home or Cards put in front of members.
Nobody is at fault for that. Velocity is the point. But velocity without visibility produces a predictable and expensive set of outcomes. Duplicated patterns. Disclosure language that drifts by vertical. Three teams solving the same comparison problem three different ways. And a design org whose institutional knowledge lives in the heads of whoever happened to be in the room.
So I stopped auditing one flow and mapped the entire product.
Building the picture
I used AI for the extraction. Every live surface, what it does, what it collects, where it sends you. Then I grouped those surfaces into flows.
The grouping is what turned an inventory into an asset. A pile of screens tells you nothing. Screens organized into flows show you the shape of the business: where journeys converge, where they fork, where we are spending design capacity twice on the same problem.
Two years ago this was not a thing one manager could produce alongside the job. The extraction cost has collapsed. Which means the scarce work is no longer the mapping. It is knowing what to ask the map.
Putting numbers on it
A map of what exists is a reference document. A map of what exists weighted by usage is a prioritization tool.
I connected analytics to the nodes. I have it on key screens today, and even partial coverage changed how the map reads. Nodes stopped being equal. A handful carry enormous volume. Others survive on nothing but inertia, still consuming maintenance and review cycles we could spend elsewhere.
Full coverage is the version I am working toward, and it forces a question I think our whole discipline should be answering: which number belongs on a screen?
Sessions is easy and tells you almost nothing. Unique members is better. Entry rate, meaning how often this screen is the first thing someone sees, would reshape how we staff. Node level drop. Time to next node. My working position is that a decision screen and a disclosure screen need different numbers on them, and that the right answer is two or three, not one. I would rather get that definition right than get the dashboard up fast.
The map corrected our mental model
Here is what the whole exercise paid for by itself.
We design as though members start at the beginning. Onboarding, then product. Our diagrams read left to right because our assumptions do.
That is not the traffic. A large share of sessions begin in the middle, arriving by deeplink from email or push. Someone lands on a refinance offer having never seen the screen we assume precedes it. They show up with none of the context we carefully built, on a screen designed to be step four.
Put every entry point on one canvas and linear design stops being defensible. Entry becomes a first class design consideration rather than an edge case, and it changes the questions I ask in review. What does this look like cold? What are we assuming they already know? Did we ever actually tell them?
Querying the surface instead of remembering it
The second unlock was pattern search across the full product.
I started with disclosures. Where they sit, how many there are, what they say, whether they read consistently. I can now answer that across five verticals in an afternoon rather than tasking five designers to go look and reconciling what comes back.
The same query runs on anything repeated. Consent language. Income collection. Empty states. Every use of urgency. Every option we preselect on a member's behalf.
That last one is the reason I care about this beyond efficiency. I have built my practice around refusing to ship dark patterns. Holding that position in a design review is straightforward. Proving it across a product at our scale is the hard version. A map with every persuasion mechanic tagged converts a stated value into a standing audit, with coordinates on every finding. That is the difference between a principle and a control.
Why it has to be visual
The shortest description I have landed on: Mobbin plus analytics, for your own product.
Designers already use Mobbin to see how other companies solve a problem, because it is visual and complete. Almost no team has that for the thing they actually own. We have Figma files organized by project and quarter. Those are a record of how work happened. They are not a picture of what exists.
I am visual. Most designers are. Most PMs are too, whatever they say in planning. A spreadsheet of screens gets skimmed once and never opened again. A map gets zoomed into, screenshotted, and pasted into Slack with a circle drawn around something. Adoption is the only metric that matters for an internal artifact, and adoption follows the format.
What I would build next
Everything so far is a snapshot. Accurate the week I made it, decaying from the moment I stop maintaining it by hand. With engineering support the map keeps itself current: nodes update as tests roll out, a new variant appears the day it goes live and sits beside control, and when the test resolves the winner becomes the node with the learning written onto it. Version history on every screen. What we tried, what happened, what we decided.
That alone would beat every design knowledge repository I have worked in. The reason I am pushing for it is what comes after.
A structured map of every live surface, grouped into flows, weighted by usage, annotated with patterns, is a machine readable model of the product. I can pass it to any model as context and hand it what no codebase or Figma library conveys: what we actually do, for whom, in what order, and how often. The map becomes the passage every AI tool we use reads first.
That changes what AI is good for here. Today it writes copy for a screen. With the map it can tell me the screen duplicates a pattern Cards shipped in March, that the disclosure placement breaks the convention we hold everywhere else, and that the preselected toggle in the proposal is the mechanic we pulled out of Insurance for cause. An eval layer for design, run before we ship rather than after a member complains.
Attach live test data and it compounds. A model holding full context on every experiment across five verticals knows what is winning, and can point to the other nodes with the same job to be done where that pattern has never been tried. Cross pollination today depends on who happens to remember what. Human memory is the bottleneck on a portfolio this size.
I am building the manual version now, because a working artifact argues better than a proposal does. Start with the picture. Everything else becomes possible on top of it.
About the author
Ashwarya Subhluxmi has 12+ years of design experience in fintech, leading teams building AI products and marketplaces used by millions of people. She writes about design leadership, AI-native product design, and what changes for design orgs when agents join the team. ADPList Top 50 UX Coach, 2026. She advises early-stage founders and is based in the San Francisco Bay Area. uxbyash.com · LinkedIn: https://www.linkedin.com/in/ashwaryasubhluxmi