Shafan Khan

A former NVIDIA engineer signed up for our waitlist. 25 minutes later, he changed how we think about engineering with AI. We’ve been building Squad by staying close to the people we’re building for. Most user calls teach you something. Some teach you a disproportionate amount. This was one of them. We were talking about how to create better specs for agents. A spec is the clear plan: what to build, why it matters, what constraints exist, and what “done” means. His answer was simple: Sometimes, to build the future, you have to look to the past. Then he pointed us to a book: *Mastering the Requirements Process.* The lesson was not “write more docs.” It was that agents don’t need more prompts. They need better instructions, boundaries, and feedback. Right now, a lot of AI development feels like brute force: More tokens. More compute. More agents. But great engineering teams don’t run on brute force. They run on clear goals, shared context, standards, reviews, and tests. Agents need the same structure. The problem is that a lot of work depends on context that was never written down. The thing someone said in a meeting. The tradeoff buried in a Slack thread. The reason a decision was made. The edge case everyone assumes is obvious. Humans can sometimes fill in those gaps. Agents usually can’t. If the context is missing, they guess. If the goal is vague, they optimize for the wrong thing. If “done” is unclear, they stop too early or keep going too long. The clearer the spec, the better the output. That connected directly to what we’re building with Squad. The value is not “AI writes code.” The value is turning messy human intent into work agents can execute and humans can review. The strongest validation? He had already built a version of this internally. Docs to specs. Specs to agents. Agents to tested work. But it was hacked together across tools. His point was simple: “I’d rather pay for a productized version of this than maintain our internal one.” That is why we keep having these conversations. You learn what is noise. You learn what is real. You learn which parts of your idea matter more than you realized. As AI gets better, do we need less process or better process?

peak engineering is watching your macbook just shit out code, letting it self-improve via layered verification and computer-use e2e before it moves onto the next task