Prompt to Product

AI Comparison Case Study

Overview

This project started as a question with no clean answer — with so many AI app builders launching at once, which one is actually worth using? Rather than read about them, I built the same app in four of them using the same single prompt and evaluated each on output quality, design decision, iteration responsiveness, and free tier usability. What started as a tool comparison became something more useful: a lesson in why the design process still matters even when AI is doing the building.

View Full PRD

Sections

Discovery

Before any of the four AI app builders were tested, the app was built once in Replit — unplanned, with no design direction, no wireframes, and no defined feature list. The intent was simply to start building and see what came out. What emerged from that first session was rough but revealing. Seeing a working version of the app, even an incomplete one, made it immediately clear what the experience actually needed — where the navigation felt missing, which interactions didn't exist yet, and what the feed was lacking without a summary view or engagement actions. Every gap that surfaced became a requirement. Every friction point in the flow became a feature to prompt for. Replit didn't define what the app should look like — it defined what the app needed to do. That distinction became the foundation for everything that followed. The feature list in this PRD is not a spec written in advance — it is a record of what was learned by building first. Replit is the control standard every builder in this evaluation is measured against, not because it is the best output, but because it is the most complete reference point for what the finished app should have.

The Builders

Before any builder was tested, the app was built first in Replit with nothing more than a problem statement. No wireframes, no feature list, no design direction. What came out of that first build wasn't polished — but it was enough to react to. Seeing the app exist in any form made it immediately clear what was messing, what felt wrong, and what interactions the experience actually needed. Every friction point, every missing feature, every UI decision that needed revisiting became a requirement. Replit wasn't part of the evaluation — it was the discovery phase that made the evaluation possible, and the standard every builder is measured against.

v0 by Vercel - Brevity

v0 produced a strong start — naming the app Brevity, establishing a two-color design language, and delivering a working two-screen flow from a single prompt. The category filter pills and card layout were production-quality from the first output. Across three iterations it handled additive requests cleanly but struggled with precision fixes — layout drift output the mobile frame, a read indicator that needed multiple passes to stick, and modal scroll behavior that required its own dedicated correction prompt. The labeled version history was the standout organizational feature, making it easy to navigate between builds. Best for designers already in the Vercel ecosystem who want a strong, minimal foundation fast.

Lovable - Quick News Digest

Lovable made a strong first impression compared to v0. Brevity had a clean and minimalistic approach, Lovable came with personality — color coded categories and emoji accents that gave the UI an intentional warmth. The feed was immediately more scannable with the category color mapping through from the filter pills to the article cards. Two iterations were enough to build out the fill feature set, and the builder responded to both prompts with minimal friction.

Feature title.

This is a feature description spanning a couple of lines.

Feature title.

This is a feature description spanning a couple of lines.

Feature title.

This is a feature description spanning a couple of lines.

Feature title.

This is a feature description spanning a couple of lines.

Key Takeaways

Key Takeaways

This project taught me more about the importance and details of the design process than it did about AI tools. Going striaght into a builder without a design direction produced something — but reacting to what came out wasn't the same as designing with intention. They discovery phase in Replit revealed that AI builders are excellent at executing a clear brief and much weaker at filling in what wasn't specified. The builders that produced the strongest outputs weren't necessarily the most technically capable — they were the ones that made the best design decisions in the absence of explicit instructions. That's a design judgement call, not an engineering one, and it's where the gap between a good builder and a great one actually lives.

This project taught me more about the importance and details of the design process than it did about AI tools. Going striaght into a builder without a design direction produced something — but reacting to what came out wasn't the same as designing with intention. They discovery phase in Replit revealed that AI builders are excellent at executing a clear brief and much weaker at filling in what wasn't specified. The builders that produced the strongest outputs weren't necessarily the most technically capable — they were the ones that made the best design decisions in the absence of explicit instructions. That's a design judgement call, not an engineering one, and it's where the gap between a good builder and a great one actually lives.

Final Verdict

Final Verdict

Feature description.

Feature description.

Feature title.

Feature title.

Feature description.

Feature description.

Replit was used during discovery phase to define the feature target and is the control standard for this evaluation.

←

trainON

↺

Return to Projects

Tellus

→