The demo looked great. Every screen, taken alone, still does. But your signup curve is flat, users bail somewhere between account creation and first value, and you can't point to a single broken feature. Here's the uncomfortable diagnosis: nothing is broken. The product just disagrees with itself on every screen. That condition has a name: design debt.
Key takeaways
Design debt (Nielsen Norman Group calls it UX debt) is the experience counterpart of tech debt — which NN/g defines as the cost of "launching faster or easier technical solutions, instead of releasing the best approach" — and left unaddressed, it compounds into user problems and costly cleanup [C1][C2].
Every prompt you send to a coding agent generates locally plausible UI with no memory of a global system. The model isn't sloppy; it's stateless. It solves this screen beautifully, then solves the next screen beautifully in a slightly different way. Inconsistency isn't a failure mode of vibe coding — it's the default output.
Practitioners see the same fingerprints everywhere. Boldare, an agency that inherits these codebases (so, yes, they profit from the problem), describes the pattern precisely: border-radius 4px on one screen, 8px on the next; the primary action color is #2D6BE4 in the nav and #2B6CE3 on the landing page; "Heading typography varies between sections because each prompt generated styles from scratch." Their summary line deserves framing: "Time saved on generation comes back as visual debt that needs to be paid down." [C11] That's qualitative practitioner evidence, not a study, but the same mechanism has been measured on the code side. GitClear, a code-analytics vendor, found block duplication up 81% from 2023 to 2026, copy/pasted code at 15.7% of changed lines in the first half of 2026, and moved-or-refactored code collapsing to 3.8% [C8]. No global memory means duplicate everything, consolidate nothing. Your UI is running the same algorithm.
There's a name for what accumulates. Nielsen Norman Group frames UX debt as the experience counterpart of tech debt, which it defines as the cost of "launching faster or easier technical solutions, instead of releasing the best approach" [C1]. As Anna Kaley of NN/g puts it: "Like tech debt, UX debt piles up over time and, if left unaddressed, leads to compounding user problems and costly cleanup efforts" [C2]. That definition was written in 2018 — before AI made the accumulation automatic. (NN/g sells UX training, for what it's worth; the concept has held up regardless.)
When a founder hears "design debt," they picture ugly. Wrong frame. The cost isn't aesthetic — it's cognitive.
Users learn a product by building a mental model of it, and consistency is the raw material of mental models. Jakob's Law (popularized through Jon Yablonski's Laws of UX) says users prefer your product to work the same way as everything else they already know [C10]. Nielsen's fourth usability heuristic makes the internal version explicit: "Users should not have to wonder whether different words, situations, or actions mean the same thing. Follow platform and industry conventions." [C5]
A vibe-coded product violates this by construction. Is "Save" on this screen the same as "Apply" on that one? Is the blue button the primary action here, or is it the outlined one like on the settings page? Each tiny question is a toll, and users pay it on every task. Most won't complain. NN/g's blunter phrasing: too much UX debt "will result in a loss of trust, traffic, and revenue," and "many customers will give you one try and then give up" [C3].
The place this shows up in your metrics is activation. Userpilot's benchmark across 62 B2B companies puts average activation at 37.5%, with the median around 37% [C9]. Userpilot sells onboarding tooling, so read the number with that in mind. Still: the baseline is already brutal. Nearly two-thirds of signups never reach first value at well-run companies. A comprehension tax on every screen pushes you below that line.
Reid Hoffman's line still circulates for good reason: "If you're not embarrassed by the first version of your product, you've launched too late" [C13]. He's right — for pre-launch. Before real users arrive, polish is procrastination and shipping ugly is a virtue.
The line moves the day real users show up. From then on, three things happen. The debt compounds, exactly as the 2018 definition predicted [C2]. The drifted patterns metastasize: every new prompt copies the nearest existing example, wrong ones included, so today's two blues become next quarter's five. And "later" quietly inflates from a cleanup sprint into a rebuild.
So ship-ugly and audit-early aren't in conflict. They're phases. And the audit isn't a redesign — it's triage, ranked by what's blocking activation, not by what offends a designer's eye.
A UX audit of a vibe-coded MVP runs four passes: a heuristic evaluation against Nielsen's ten usability heuristics, an activation-funnel walk from signup to first value, a design-system triage that picks one canonical version of every duplicated component, and a fix list ranked by activation impact.
Here's how we'd run each pass on a live vibe-coded product.
Pass 1 — heuristic evaluation. Walk the whole product against Nielsen's ten usability heuristics. These were developed with Rolf Molich in 1990, refined in 1994, and have "remained relevant and unchanged since 1994" [C4]. They're about human perception, not visual fashion, which is why they survived three decades of trends. One warning from the same research tradition: don't solo this. In Nielsen's 1994 findings, averaged over six projects, single evaluators found only 35 percent of the usability problems; the standard is three to five people evaluating independently [C6][C7]. "I clicked around and it seemed fine" is a 35% audit, run by the one person on earth least able to see the product fresh.
Pass 2 — the activation-funnel walk. Signup to first value, step by step, counting every stall, ambiguity, and dead end, then judged against that 37.5% benchmark [C9]. This pass is what keeps the audit honest: findings that don't sit on the activation path get demoted, however ugly they are.
Pass 3 — design-system triage. Inventory every component in the product. Find the near-duplicates: the two blues, the three button styles, the four card layouts. Pick one canonical version of each and write it down. This pass matters more than it sounds, because it's how a vibe-coded product acquires the design system it never had. More on why that matters in a moment.
Pass 4 — the prioritized fix list. Everything found, ranked by activation impact. A good fix list says out loud that most findings aren't worth fixing this quarter. If your audit deliverable is 200 items with no ranking, you paid for a complaint, not a plan.
The strongest counterargument is that tooling will fix this upstream. Figma's case for design-system-aware generation is real: give agents your system as context (via MCP) and they "[r]euse existing components and patterns: [r]educing duplication and inconsistency" [C12]. Ana Boyer's analogy is the right one — generating without that context is "like asking a new engineer to start shipping code before onboarding" [C12]. Figma sells this tooling, but the logic stands on its own; we ship with AI daily and system-as-context genuinely helps.
Notice the presumption, though: it works if you have a design system. The vibe-coded cohort is precisely who doesn't — that's the whole diagnosis. Which makes the audit's triage pass the bootstrap: run it once and you get the canonical component set that makes your AI tooling dramatically better from then on. The tooling and the audit aren't rivals. One is the onboarding doc; the other is how the doc gets written.
Who should run it? Someone who does product design, not graphics. You need the person who reads flows and mental models — where users stall, which words collide — not someone who'll restyle your buttons. In DevOD terms that's a UX engineer: an engineer whose discipline is product and UX design work, not decoration. Your options (freelancer, agency, audit shop, subscription) trade off differently enough that we wrote a separate buying guide on hiring UX help, plus a breakdown of what a UX designer costs to hire in 2026. Whatever route you pick, an audit is close to the perfect trial task: fixed scope, concrete deliverable, easy to judge.
Skip the audit for true prototypes, internal tools with a captive audience, and anything without real users yet. The line is the same one we draw for testing AI-generated code: real users plus real data. Before that line, Hoffman is right and your embarrassment is a feature. After it, the debt is billing you daily whether or not you've looked at the invoice.
Every prompt ships UI. Someone has to ship the system.
A UX audit happens to be the exact shape of DevOD's one-task Proof of Quality: a 15-minute fit call, then hand the audit to a UX engineer as the trial task, get back the prioritized fix list, and judge the work before you subscribe. If it's good, that UX engineer sits on one subscription alongside your dev and QA engineers — add or change roles any month, cancel any time, $3,495/mo for a single stream, with a 3-day task cycle and daily async updates.
The model won't notice the two blues. Your users already have.
What is design debt? Design debt (Nielsen Norman Group calls it UX debt) is the experience counterpart of tech debt, which NN/g defines as the cost of "launching faster or easier technical solutions, instead of releasing the best approach." Left unaddressed, it compounds into user problems and costly cleanup efforts [C1][C2].
What does a UX audit include? Four passes: a heuristic evaluation against Nielsen's ten usability heuristics, an activation-funnel walk from signup to first value, a design-system triage that picks one canonical version of every duplicated component, and a prioritized fix list ranked by activation impact. It's triage, not a redesign.
Why is AI-generated UI inconsistent? Coding agents are stateless: each prompt generates locally plausible UI with no memory of a global system, so styles drift screen to screen — two near-identical blues, three button styles [C11]. GitClear measured the same mechanism in code: block duplication up 81% from 2023 to 2026 [C8].
When should you do a UX audit on an MVP? Once the product has real users and real data. Before that line, shipping ugly is a virtue and polish is procrastination. After it, the debt compounds daily and every new prompt copies the drifted patterns forward. Skip audits for true prototypes and captive-audience internal tools.
How many people should run a heuristic evaluation? Three to five evaluators, working independently. In Nielsen's 1994 findings, averaged over six projects, a single evaluator found only 35 percent of the usability problems [C6][C7].
