Beyond Vibe-Coding: A Strategic Framework for AI in Design

Published in 2025 by Eric

Nearly every day I come across some new riff on how vibe-coding with AI can accelerate product development. This is exciting stuff with lots of potential.

However, the folks who’ll benefit the most from these early ideas are lean, nimble entities like early stage products and startups. So what about everyone else?

Vibe-coding serves merely a small slice of the overall product design process. In fact, most instances I’ve seen of AI providing value in the enterprise are for things other than vibe-coding. This is especially true for B2B enterprise, where complex webs of teams must coordinate and negotiate realistic steps towards delivery.

Below, I’ll share a framework for how I’m thinking about the possibilities of working smarter with AI as designers. I’ll share a framework, along with some specific methods I’ve applied in my own process.

A Familiar Model

Let’s start with something familiar: the Double Diamond. We can use this to identify new methods where AI might support us at each step in the process.

Vibe-coding, the most familiar process, leads us to think of AI’s potential as an “accelerator” that blurs roles downstream between design and engineering. And indeed, it may this in some contexts. If we zoom out and look across the entire process, however, the opportunity here is about much more than speed or effort.

Our potential for AI in design is about stronger strategic impact. More trust with our partners and customers, helping all of them succeed at more elite levels. Put another way: AI has the potential to increase the speed that we reach higher quality decisions, strengthening trust with product partners and customers.

For design, especially at product-led companies, this unlocks what we’ve always dreamed of: Supporting and influencing product decisions.

Let’s discuss some specific methods.

Strategy Refinement

🔷 Discovery

Better Roadmap Prioritization Systems

For many companies, prioritizing features remains a bit of an art form. However, AI can help you turn that into an objective, quantitative system, tailored to your company’s needs. It can not only help you identifying the right measurement and tracking framework, whether it’s a version of RICE analysis backed by KANO research.

Perhaps the biggest benefit: The prompting exercise itself forces a conversation between you and your partners to articulate, align, and deeply understand the factors that drive your roadmap.

Community Feedback Aggregation

Online critics are notoriously some of the harshest. For designers, that harsh criticism of our products is a valuable—but often untapped—research input.

Keeping afloat of these conversations has traditionally required manual searches, subscribing to key industry voices, and keeping a close eye across many different surfaces. Because of this extra lift, it’s often neglected.

With a good prompt and automated reports, designers can now get frequent reports on the latest chatter around our products. A goldmine of user feedback waiting to be unlocked.

Tapping Into Past Research

Oftentimes, great research happens, and then we move on to the next topic. However, most research, especially discovery, has a lifespan of value that’s often untapped. When you have 10 studies to juggle, and some of them a few years old, it’s easy to have a recency bias towards the latest ones.

Using recorded videos, anonymized transcripts, or just synthesis decks, we can use AI to collect data across our entire research catalog to test our hypotheses. This is uniquely valuable in that often users provide feedback on topics we’re not even testing for in a given study, which I call peripheral insights.

Synthetic User Testing

Ok, Gemini came up with a few ideas in here too, and here’s the first. I haven’t tried it, but here you go:

Concept: Instead of general feedback synthesis, create persistent “Reflective Agents” (based on the Stanford notes) representing your key personas.
Action: Feed these agents every PRD and Release Note. Before you build, ask the “Council”: “Based on your history of frustrations with our app, how does this new feature change your sentiment?”
Value: Predicts “feature fatigue” and long-term trust issues before they happen.

🔷 Definition

Triaging PRDs

Product’s primary tool, traditionally, has been the written word. Often pages of words. This has worked well for many teams, especially when there’s a coherent PRD structure.

When you’re juggling half a dozen or more PRDs, it’s helpful apply different lenses to them. With tools like Google NotebookLM, not only can you centralize and ask questions, but you can convert them into mind maps and engaging videos.

Vibe-Coded Prototypes (Strategy)

There are two types of vibe-coded prototypes: strategic and executional. Vibe-coded PRDs provided by Product are most often strategic.

We can think of these as “napkin sketches“, visual ways of getting an idea across, and getting rapid alignment across many teams. When a designer receives one of these, our responsibility is to zoom out and map out the system and the intent communicated with the sketch. When we zoom back in, we’ll often find ourselves at a completely different solution.

We should be very grateful to any product partners who use this tool, as napkin sketches are catalysts for cross-team alignment and feedback.

Quantitative Research Dashboards (Internal Tool)

Honestly, which designer has time to learn and configure BI tools? I’ve spent a good amount of time in Tableau, but still nowhere near proficient. To really make insights actionable, you need to transcend what BI tools offer, even for advanced users.

My big learning here was iterating on a Tableau dashboard as a vibe-coded Google App Scripts dashboard. It was completely customized to suit our needs. This vibe-coded dashboard pulled from heuristics data in Google Sheets, display items in a card format, and link to specific rows within the source spreadsheet.

Workshop Planning

Designers uniquely excel at bringing diverse stakeholders together—but this takes a lot of planning and coordination. Finding ways to break the ice, engage partners, schedule, allocate teams. It’s an art and a science, and very time consuming.

Given a set of goals and guardrails, the right prompts can be hugely helpful here for planning, brainstorming, and even synthesizing the results.

“Living” PRDs

Another one (which I haven’t tried), brought to you by Gemini:

Concept: Static PRDs become old-school.
Action: Implement a RAG (Retrieval Augmented Generation) system where your AI doesn’t just summarize Slack, but updates the PRD based on the Slack decision.
Tooling: Connect Slack -> Zapier -> Notion/Linear AI. If a decision is made in Slack: “We are cutting feature X,” the AI should comment on the PRD: “Feature X deprecated per discussion on [Date].”

Execution Refinement

🔷 Develop

Vibe-Coded Prototypes (Executional)

If you’re lucky enough to have a fancy license, you can hook up a Figma server to share files directly with Cursor (or whatever tool)—creating a live, rapid prototype. Alternatively, you can paste images of screens into Cursor, and let it work its magic. Or paste a single high res screenflow, so it can weave them together into a quick demo.

This is good for getting a quick “feel” of a design your thinking through. Partners gain confidence in the decision, engineers get a better sense of the system, and you yourself will know where to iron out any quirks.

Edge Cases & Errors

Sometimes error and edge cases can hit us last-minute during delivery. Getting ahead of engineers on what errors, empty states, and other scenarios might need to be accounted for, can help us plan out better guardrails, errors, and warning systems, and have a more purposeful experience when things don’t go perfectly for users.

Dummy Data

I’ve tried numerous things to make hi-fi mockups feel more realistic: Mockaroo and Fadagen 2000 most notably. For data visualization, you can use AI to mock up realistic charts without spending countless time resizing individual bars on a stacked bar chart, for instance. The way we do that is going to evolve, but this is a key area I plan to continue experimenting with.

The “Verifier” Agent (The Rabbit Poop Detector)

Another one (which I haven’t tried), brought to you by Gemini:

Concept: Acknowledge that AI code/design is probabilistic and prone to errors.
Action: Build a separate “Critic” agent that only reviews output against your Design System guidelines and Accessibility rules. It does not generate; it only rejects.
Value: Reduces the “Seam” friction where developers have to clean up messy AI code.

🔷 Deliver

Given my focus on strategy, this space is blank. QA and relationship with engineering is such a hands-on, gritty thing. To be sure, purposefully-crafted design files for intended audience is something that will evolve, and AI will increasingly become a part of it.

But the real value of AI to our process lies further upstream.

A Note on Agentic Maturity

We can think about each phase of the diamond as having three levels of automation. In my opinion, many things that use the word “agent” are not truly agents, but rather glorified prompts, supported by structured or unstructured data. Here’s another way to think of it:

Level 1: Prompts

The most basic system: Type and paste/upload files into a prompt, and it answers. Mold that ball of clay until it’s where you need it.

Level 2: Prompts with Data Retrieval

Instead of pasting/uploading, tools at this level can pull from live data sources. For instance, if someone suggests a change

Level 3: Agents

This level is closer to how humans function.

Level 4: Multi-Agent Collaboration

Multiple agents interact with each other in a web of feedback loops. An example of this is “The Council.” a workflow where one agent generates a vibe-coded UI idea, a second agent reviews it for UX quality, and a third agent drafts the copy. You only review the final output.

Where should we be, right now?

Each method described can be achieved at any of these levels. Personally, I’m still at a nascent version of Level 2, but ideating my way towards Level 3.

To unlock Level 3, some ecosystem barriers need to be unlocked, such as cross-service integration and permissions (ie, between Slack, Google Sheets, Figma, etc). This is tricky for most companies due to security concerns, but more feasible for personal uses.

As for Level 4? That’s an exciting thought exercise, and some examples do exist out there. But I haven’t seen it applied in practice—yet.

Summary

Vibe-coding gets all of the attention these days, but I hope this essay demonstrates that our opportunity is much greater than that.

For so long, product design has been hindered from unlocking its full value to partners and customers. More deeply understanding users, steering strategy, influencing roadmap, and driving top-quality decisions.

Now is our time to shine!

AI is evolving rapidly, so consider this is a living essay. I plan to add to it as I refine my process.