Why We Launched a New Website in 30 Days
Ati Motors approached BFT to help them build a new website as Ati Robotics, and they didn’t want a traditional website build. They came with a bold challenge: Build the new site with Claude. In 30 days.
We knew it would disrupt our workday and upend our process. But we said yes (if you’ve been around for a while, you know BFT likes a challenge!). We were ready to see what we’d learn.
Over the next month, we leaned into Claude hard. We workshopped live on calls, tightened prompts, tested ideas fast, hit token limits, and rebuilt what didn’t work. We celebrated the wins, owned the misses, and kept moving.
It wasn’t all pretty.
Lesson 1: What Claude Cowork Actually Looks Like in a Real Project
First, new tools like this have a large learning curve. It’s easy to get started using Claude the wrong way, especially if you think of Claude like an advanced chatbot instead of a coworker to prompt very specifically with brand guidelines and specific examples of work. To enable Claude Desktop and Cowork to work alongside us, we had to set the parameters. You can use instructions to teach Claude how to work each time, including markdown files and notes about yourself.
Claude stopped feeling like a tool we “visited” and started feeling like a teammate for collaboration we kept nearby. A little creepy, but true. (I heard a coworker refer to Claude as “he” the other day.)
Several services, access levels, and various tools needed to be spun up to get the right connections in place for Claude Desktop. For example: We had to find free web hosting and automate connections to GIT deployment, setups, merging, branches, and commit protocols.
“
Taking the time to set things up correctly was invaluable, it helped us make sure that Claude’s outputs were consistent and we all understood what handoffs and coworking/collaboration looked like across multiple SMEs.”
– Josh Stauffer, CEO and President
Lesson 2: How We Thought We Would Work With Claude, and the Reality
One of the first things we learned the hard way: We thought we could run scheduled workshops so each subject matter expert could give insight while we iterated. We put them on the calendar, thinking this would help some of the “sprint” work.
That doesn’t translate cleanly when Claude is in the mix.
When only one person can actively drive Claude Cowork at a time, momentum gets lumpy—work can pause while Claude runs big changes or thinks through a complex task. Processing speed isn’t automatic if it’s extended thinking. More than once, we found ourselves on collaborative calls half‑multitasking, half‑waiting for the designated driver to finish a pass before we could react and iterate.
Claude can hold context, but it can lull you into thinking you don’t need guardrails. We did.
Pretty quickly, we realized we had to level up our documentation—clear reference files, consistent decision logs, and “here’s the source of truth” materials inside the project. That discipline kept the work from drifting and made it possible to hand off ownership between SMEs without losing time re-explaining what had already been decided.
Even with GitHub access and a full team around it, building with Claude still felt more like a solo build than a true multiplayer exercise. You can absolutely collaborate—reviewing outputs, refining prompts, testing together—but the hands-on creation tends to bottleneck around whoever is at the keyboard with Claude in front of them. It’s a different kind of collaboration, and we had to adjust our expectations (and our workflow) to match.
We also thought it would be a lot cleaner to share files within Claude Cowork. While there are Projects folders and Artifacts available, if you’re not the creator, it’s super hard to share Chats and context within the team. We all said we think Claude is likely working on a more enterprise-grade beta right now, and we’re just not cool enough to be testing it yet.
Lesson 3: Using Claude Cowork and Context Switching on Tasks
Once Claude became part of the day-to-day build, we noticed something unexpected: it subtly encourages you to juggle more than you should. You can tee up a task, wait for an output, review it, then bounce to the next thing while you’re “in motion.”
In theory, that sounds efficient.
In practice, hopping between two or three threads at once came with a real mental tax. The team called it out directly—we felt more intellectually exhausted than we normally do on a sprint, even when the hours looked the same on paper.
The other side of that coin is the learning curve: it’s steep, fast, and kind of whiplash-inducing.
One minute you’re thrilled because Claude helped you unblock something in seconds; the next, you’re frustrated because the same approach fails when the context shifts. You find yourself re-explaining the prompt, pointing out when Claude drifted and wasn’t consistent with its own outputs, and editing things you didn’t think you’d have to touch. We were using Claude daily, and the number of times the software updated was insane. We had a stumble when a bad update caused us to lose a day of work. Claude Design also went into beta during this launch period.
“
This kind of work puts your critical thinking on the line in a way traditional design doesn’t. The mindset shift is real, and it’s not a small one.”
– Daniel Gallegos, Senior Designer
Stay Tuned for a Full Claude and Ati Robotics Case Study
Follow our social feed for the full Ati case study. We’ll share the metrics, break down the workflow, and show exactly how we built and handed off the Ati Robotics site.
What about you? Have you had your own experiences with Claude? Reach out to your favorite BFT team member to see more examples of what we’re working on right now with AI tools like Claude, Gemini, ChatGPT, and Figma Make.
