Abstract glowing root system beneath a floating rectangular screen, representing written design intent supporting the visible interface.
June 10, 2026

The Sentence Beneath the Screen

How AI exposed product design’s clarity problem

A few months ago I was early on a dashboard project, using AI to help think through the problem before opening Figma. I typed out what I knew: the users, the data, the rough business goal. The output came back generic, a collection of familiar patterns and best practices that could have been pulled from any product design article.

So I added more context. The output improved, but only slightly. I tightened the constraints. I clarified what success looked like. I spent more time describing what the user was trying to accomplish rather than what features they needed. A few iterations in, the responses became noticeably better, specific enough to argue with, pointed enough to explore.

My first instinct was to credit the prompting. Then I looked back at the original version.

The AI hadn’t improved. My understanding of the problem had. Each revision wasn’t a better prompt. It was a clearer thought. The final version contained something the first version didn’t: a position. It said what the dashboard needed to help users do, why that mattered, and what success might look like. The earlier prompts described a product. The later ones described an outcome.

That clarity hadn’t existed before I wrote it down. Not in the brief, not in the kickoff notes, not even in my head in a form I could have defended in a design review. The machine forced it out, and what it exposed wasn’t a prompting weakness. It was a thinking gap that had been there from day one.

The more I’ve worked with AI, the more I’ve become convinced this is one of the most important lessons hiding beneath the current conversation about prompting. AI didn’t create a clarity problem in product design. It exposed one that was already there.

The mockup hid the thinking

For most of my career, product design has been remarkably forgiving when it comes to ambiguity.

A mockup can look resolved long before the thinking underneath it is. Screens are persuasive. A polished interface creates the impression that a decision has been made, even when the reasoning behind it remains fuzzy. Teams spend weeks refining solutions before they can clearly explain the problem those solutions are meant to solve. Discussions drift toward hierarchy, interactions, and visual polish while more fundamental questions stay unanswered.

Writing doesn’t offer the same cover.

Write “we believe simplifying onboarding from six steps to three will improve completion because users abandon the process during identity verification” and every part of that statement becomes visible. The assumption is visible. The hypothesis is visible. The evidence — or absence of it — is visible. Someone can challenge it. That’s exactly why writing matters: it exposes uncertainty while it’s still cheap to address.

AI amplifies this effect. The moment you ask a machine to execute against your thinking, gaps that might have remained hidden for weeks become obvious in minutes. The machine doesn’t care how polished your mockup is. It can only execute the clarity you’ve provided.

The framing was wrong

On a recent project, I asked AI to help explore concepts for a dashboard that would help operations teams manage risk. The outputs were perfectly reasonable, alert systems, escalation workflows, risk scores, notification patterns. None of them felt right.

After a while I realised the issue wasn’t the quality of the concepts. It was the framing. The real goal wasn’t helping users identify risk. It was helping them understand confidence.

Those ideas sound similar, but they lead to very different products. Risk identification pushes toward alerts, warnings, and escalation. Confidence pushes toward transparency, context, and explanation. One centres on detecting problems. The other centres on helping users understand uncertainty. The moment I rewrote the brief around confidence instead of risk, the outputs changed. The metrics changed. The information hierarchy changed. The suggested interactions changed.

My articulation of the problem was wrong, and the machine had faithfully executed that wrongness at speed.

This is the part of AI most discussions skip. The technology acts as a thinking test, not because it understands the problem, but because it executes whatever understanding you’ve provided. Vague thinking produces vague output. Contradictory thinking produces contradictory output. Clear thinking produces something worth arguing with. Output quality is capped by thinking quality, and AI makes that ceiling visible faster than anything else in the process.

Minimal stone slab floating above clouds, supported by a glowing foundation block, representing an executable brief guiding product decisions.

The brief is becoming executable

In my last article, I argued that briefs should become systems, not documents read once at kickoff and forgotten, but living sources of truth that guide decisions throughout a project. This piece is really the missing piece of that argument. A system is only as rigorous as the thinking it contains.

Human teams are good at compensating for unclear thinking. Designers infer intent. Product managers provide context. Engineers fill gaps. Everyone carries part of the understanding in their heads, and the project moves forward on a foundation of shared assumptions. That ability is one of the reasons ambiguity survived in design practice for so long.

AI doesn’t participate in those assumptions. It only sees what has been articulated, not what you meant, not what stakeholders assumed, not what the team informally agreed to. It executes against what exists in writing, which means weaknesses that once stayed invisible become obvious fast. The machine cannot reconstruct context that was never written down.

That’s why the discovery document, the design rationale, the hypothesis, and the success criteria are no longer just records of intent. They function as instructions. They guide decisions across teams, across time, and across AI systems. The brief is becoming executable, and vague briefs are producing vague products faster than ever before.

The write-first teams aren’t winning because they produce more documents. They’re winning because they’ve written clearer operating instructions for everyone involved in the work, including the machines.

One reasonable objection to write-first design has always been time. Writing is hard. Good writing is harder. Most teams were already stretched producing the work itself. AI weakens that argument considerably. The friction of drafting has collapsed. Structuring, rewriting, and refining are cheaper than they’ve ever been. What remains is the hard part: deciding what you actually think.

Clarity is the scarce resource

For most of my career, production was the bottleneck. Exploring ten directions instead of three meant someone had to spend the time creating them. Every additional prototype, concept, or experiment carried a real cost attached to it.

That constraint is disappearing. You can generate concepts, prototypes, and code faster than you could have imagined five years ago.

The harder problem is deciding which direction is worth exploring in the first place. Many teams still struggle to articulate exactly what they’re trying to achieve, not because they lack talent, but because clarity has always been easier to postpone than execution. Production is no longer the scarce resource. The bottleneck has moved upstream, toward judgement, toward reasoning, toward the ability to commit to a position before a pixel exists.

The designers who thrive won’t be the ones who generate the most artefacts. They’ll be the ones whose thinking is clear enough that both humans and machines can act on it.

The most important thing AI has taught me isn’t how to write better prompts. It’s how often I wasn’t thinking clearly enough before I opened the tool. For years, the industry could hide ambiguity inside beautiful work. AI didn’t create that problem. It simply removed our ability to ignore it.

The prompt was never the problem. The sentence beneath it was.