Build #02 of Mate & Build didn't have a single climax moment. It had three. And that makes it the hardest one to tell — and also one of the most important.
"I have three things that are eating my time"
That's how Lupa started.
Luciana Panio works at a bank, Buenos Aires. Brand Expert at Fima. She came to the session with a list of processes that had been stealing hours from her: generating images for social without depending on the design team, putting together competitive-analysis decks from scratch every time, and proposing concrete changes to the internal website in a way the technical team could see and understand. Not a verbal opinion. Something navigable.
Three different tasks. Unrelated to each other. Each one with its own broken flow.
Usually at Mate & Build — the weekly AI problem-solving sessions Patricio Iturraspe runs — we work on a single problem. We grab it, break it apart, solve it in depth. This time was different. And the decision to make it different was the most important one of the whole session.
The scoping decision: why not go deep on a single one?
The obvious question was: do we pick one and solve it well, or do we go after all three?
The answer came fast once we looked at the three problems together. They had something in common that wasn't obvious at first: all three were bounded processes. None of them required complex architectures or system integrations. Each one had a clear input, a defined output, and a specific friction point that could be unblocked with the right tool.
Going deep on a single one would have been "tidier." But it would have left two real problems unsolved that Lupa also had — not because they were hard, but because solving something that "already kind of works" always gets pushed to later.
The decision was conscious: aim for a solid structural base on all three, not deep on just one. Three different flows, three different tools, three separate wins. And also — this matters — three different limits.
Problem 1: images for social without going through the design cycle
The task was specific: generate images for internal communication and social without entering the cycle of asking design → wait → request changes → wait again.
The goal wasn't "use AI to make images." It was getting to a finished image on the first try — input prompt, output image, no iterating.
We found the flow with Gemini and Nano Banana Pro. The key was packing enough context into the prompt: tone, the exact text that goes in the image, visual reference, desired format. When all of that is in the prompt, Gemini doesn't have to guess anything. The result comes out one-shot.
The limit showed up on its own, without us looking for it: images generated with the free version of Gemini come out with the Gemini logo in the corner. For professional use — publishing on LinkedIn, institutional communication — you need the premium plan. It's not a technical problem. It's a budget decision. The flow is solved; the remaining friction isn't about the tool.
Problem 2: research the market and put the deck together without starting from zero
Competitive analysis was the process eating most of her time. Researching the market, organizing the information, putting together a coherent presentation that on top of everything respects the company's visual identity. All of that, every time, from scratch.
We attacked it in two steps.
First, deep research with Gemini. We asked it to map the competitive landscape of the financial sector. Not a Wikipedia summary — a structured analysis with enough depth to build a deck off of. It took a few minutes. The output was solid.
Then, Manus AI. We handed it the bank's brand book and the research output, and asked it to put together the presentation.
What came out had the right colors, the right typography, the right style. It wasn't "something kind of like the visual identity" — it was the visual identity, applied directly on top of the generated content. That moment had a real impact: seeing that the tool not only handled the content but also correctly applied the company's visual rules.
The limitation is clear and worth saying: the output is only as good as the input. With the general information we gave it, the result was a solid starting point for the deck — useful, structured, presentable. If you give it detailed data per competitor, with specific numbers and differentiators, the presentation becomes much more robust. The difference between "starting point" and "deck ready to ship" is how much context you give the tool. But you no longer start from zero.
Problem 3: the website redesign that ended up deployed on Vercel
The third front was the most ambitious. And the one that ended best.
The task was proposing CRO changes to the bank's website — CTAs, layout, visual hierarchy, user flow — in a way the internal dev team could see and understand. Not a verbal comment in a meeting. Something visual, concrete, navigable.
Here we used Claude Code with several skills, and the most important one (in our case) was the UI/UX Max Pro skill (available on GitHub: nextlevelbuilder/ui-ux-pro-max-skill). The process was in two stages. First we duplicated the full site. Then the skill analyzed the structure — identified which CTAs were poorly placed, which visual hierarchy didn't help conversion, which flow elements created friction — and generated a complete redesign with our adjustments and conversion-oriented requests. All built on top of the real site code, not a mockup.
We deployed the result on Vercel.
Lupa has a link. She can open that link, navigate the redesigned site, and share it with the technical team. It's not a PDF with little arrows. It's not a slide deck with screenshots and comments. It's the bank's website running with the changes applied — the minimum that actually solves the problem of communicating the proposal.
That was the moment of biggest impact in the session. Seeing the redesign working in the browser changed Lupa's face.
The limitation also worth telling: we worked with the visible HTML code of the site, not with full repository access or the components folder. With full source code access, it would be much simpler to iterate on specific components and adjust with precision. What we built is a functional visual prototype — enough to show direction and drive the internal conversation, not to replace the current site directly.
What stayed on each front — and what's still missing
Three different flows, three concrete results in one afternoon.
For image generation: the process is solved. A well-built prompt with full context produces a one-shot result in Gemini. The only pending decision is whether the budget justifies the premium plan to remove the logo.
For competitive analysis: the tool works. The pipeline of deep research + Manus AI + brand book produces decks with real visual identity. The next iteration is feeding it more granular data per competitor so the output goes from "solid starting point" to "almost ready to ship."
For the website redesign: Lupa has a link deployed on Vercel with the proposed CRO changes. The next step is sharing it with the technical team and starting the conversation about what gets prioritized. With full repo access, the iteration would be much more precise.
What stayed with me most from this session wasn't any of the tools in particular. It was that the three processes had existed for months and no one had found the time to ask whether there was a better way. Not because they were complicated. But because all three "kind of worked," and that was enough not to attack them.
"It wasn't a static mockup or a PDF with little arrows. It was the website running with the changes applied — the minimum that actually solves the problem of showing the direction."
Takeaways
- Scoping is a technical decision. Choosing "go after several" instead of "go deep on one" requires analyzing whether the problems are bounded and decoupled. If they are, they can be solved in parallel without losing quality on any of them.
- The one-shot prompt isn't magic — it's context. In image generation with Gemini, the difference between iterating five times and nailing it on the first try is how much information you put in the initial prompt: tone, text, visual reference, format.
- Visual identity can be automated if you have the brand book. Manus AI with the brand book as input produces decks with the real visual identity — not "something similar." The limit isn't the tool but the quality of the content data.
- The minimum that actually solves vs. the perfect that never ships. The website redesign deployed on Vercel isn't the final solution. It's enough to move the internal conversation — and that's exactly what was needed.
- Tool limits are data, not failures. The free Gemini logo, Manus AI's dependency on input data, the lack of full repo access — all three are real limits Lupa takes home as useful information to decide what to prioritize next.
Concepts applied here
- AI for marketing (in Spanish) — how marketing professionals at companies solve their own problems without waiting for IT.
- Context engineering (in Spanish) — why the brand book + client data as context define the output quality.
- Folder structure (in Spanish) — the structure that lets the UI/UX Max Pro skill read the real site code.