My First FDE Experience - What Broke When We Replaced Squarespace With Claude
This is the first in a series of field notes about what worked, what broke, and what I learned while practicing forward-deployed engineering inside my own organization.
“Forward-deployed engineering” has gained renewed prominence in the technology sector as major AI labs like OpenAI(launched a deployment company with B&C and McKinsey) and Anthropics(partnered with Blackstone, Hellman & Friedman, and Goldman Sachs) seek to pair frontier models with hands-on enterprise implementation. Feeling inspired, I ran a small FDE experiment with a couple of internal non-technical teams. I got a firsthand look at how cutting-edge technology is changing the workplace and how people respond to and use new AI tools.
AI moved execution away from the product surface.
AI increased execution leverage but initially reduced users’ sense of control. Adoption improved only after the workflow restored a concrete way to inspect and approve the result.
The experiment was a simple “lift-tinker-shift” migration from a Squarespace website to a self-hosted static site. I wired up Claude Tag in Slack with a few instructions, a GitHub repo, and some design and marketing guidelines.
I did the first few runs and demoed what was possible with Claude Tag and assumed the marketing team could take over and I could simply monitor on the sideline.
And of course that didn’t go as planned; the marketing team quickly surfaced their concerns.
Editing through chat felt less intuitive than clicking directly on the website in Squarespace.
Managing email notifications for form responses and web page redirects does not have an interface and poses a real challenge for the marketing team to understand what’s happening behind the scenes. They had to rely on engineering for assistance.
What broke
Squarespace had collapsed intent, action, and feedback into a single WYSIWYG interface. Users clicked on the page, made a change, and saw the result immediately. The editing surface was also the finished product.
In comparison, Claude Tag in Slack (or any other AI workflow) separated those activities. Users described an outcome in Slack, an agent changed files elsewhere, and the result returned later. Although the AI workflow alleviated the user from hundreds of clicks and managing pixels, padding, and margin, the user could see less of the work while it was happening.
Why it wasn’t a problem for engineering
Software engineering experienced the same shift, but with a longer runway. Cursor started with Cursor Tab, which suggests code and next edits inside a familiar IDE. While AI changed how developers wrote code, the editor remained the center of the experience. With foundational model advancement, that center of gravity is now moving. The cursor is now primarily in the Agents Window, which orchestrates agents to run multiple sessions in parallel. The IDE remains available but increasingly acts as a surface for low-frequency inspection rather than the only place where work happens.
Early adopters (mostly software engineers) therefore experienced the transition in stages: from code completion to collaboration to delegated execution. And these stages were gated by model capabilities. Adopters today are often asked to make that transition all at once; therefore, the change feels more sudden and abrupt.
Our response: restore evidence, not implementation.
Successful AI adoption requires creating the missing transition period deliberately—preserving familiar feedback surfaces, introducing delegation incrementally, and providing hands-on support until the new way of working becomes natural.
I loved Cursor’s “demos over diffs“ philosophy, and I quickly implemented our own version when it first dropped. So naturally I drew on Cursor’s expertise and guidance and looked for a solution that would allow the marketing team to see and interact with Claude’s work easily in real time.
We chose Vercel.
Each change produced a working preview site on Vercel.
Vercel already has in-page on-element comments to facilitate collaboration and feedback.
Those comments are cycled back to Claude for step improvements.
This did not restore every control. Redirects remained engineering-owned because they affected infrastructure outside the website’s editing surface. But it repaired the highest-frequency gap: users could once again see the product, point to a problem, request a revision, and decide whether the result was ready.
Takeaway: replace a familiar control surface with a feedback surface that preserves the user’s ability to inspect, annotate, and reject the result.