When a team builds a working prototype before it writes the requirements, someone still has to write down what engineers will build and why. Two companies described that step this week, and both treat the written account as part of the delivered work.
Microsoft (4 Oct · 01) runs Foundry Labs, a website where people can find AI models, tools and research projects from teams across the company. The site was easy to browse with 20 projects and became hard to search as the catalogue grew. Principal product manager Saumil Shrivastava assembled a team of three, and technical program manager Gulsimo Osimi first wrote a 32-page specification. According to Mark Armstrong's account on Microsoft's Inside Track blog, the document proved a poor way to share the plan, so Osimi built an interactive prototype with AI tools. She then revised the specification and the prototype against each other, and software engineer Pratyansh Agrawal turned the prototype into the production site in about three months. Osimi put the limit plainly: "Developing a prototype is one thing, but making it a product is different."
Salesforce (4 Oct · 02) makes customer-relationship software and maintains several design systems across its products. Alan Weibel, a UX architect there, describes an internal pipeline called Experience Operating System that takes a design problem from an idea to an engineering handoff. A team can start with a brief, a requirements document, a Figma file or a working prototype, and the pipeline writes the earlier documents that were skipped. The handoff includes a specification covering layout, every interface state and keyboard behaviour, together with the research, feedback and reasons behind each decision. A person still decides which review comments become rules that future AI agents follow. This is Salesforce describing its own internal system.
Both teams let the prototype lead and then produce the written specification alongside it rather than after it. At Microsoft the person who built the prototype kept the specification current by hand, so the builder remained the only translator. At Salesforce the pipeline drafts the specification and its reasons, and a person decides what becomes a rule for the next AI agent. The view that this removes the builder as the only translator is an inference, not the author's claim.
Both accounts come from companies describing their own teams and tools, and neither reports how often the written specification disagreed with what shipped. An internal website built in three months and a vendor's design pipeline may say little about a team shipping a product to customers in several markets.
A three-person Microsoft team rebuilt its Foundry Labs site by prototyping with AI tools first, then revising a 32-page specification against the prototype until engineering took over.
Salesforce's design team built a pipeline in which AI agents produce designs from any starting point and hand engineers a specification that records the reason for every decision.