
Aggregated wisdom from YC founders on product building, user research, and the startup journey—organized by topic.




























Join hundreds of companies using Human Behavior to turn user sessions into actionable insights. Set up in 90 seconds.
Real insights from YC founders, organized by topic. Filter by founder, company, batch, or industry to find the perspectives that matter to you.
“We try to max Claude Code as much as possible. I want a backlog the agent can run in a loop overnight so we wake up and review five PRs every morning. We’re not there yet, but that’s the goal.”
“We use Cursor build mode with a plan doc—answer all the questions first: how are you parsing, chunking, retrieving, where does it live, is it secure? Once that’s clear, we say, “Okay, go build it.” We audit the code, but not a lot. We still look at the code so when something breaks, we can immediately fix it.”
“For our prod pipeline, it’s all handwritten. Any velocity gains would be eaten by risk of breakage, and that’s something I never want. For front end and utilities that aren’t critical to our live bidding pipeline, a lot of it is AI‑assisted.”
“What I love with Claude is sub‑agents. We’ve programmatically trained them on our product, security, compliance, auth—so anyone can use Claude and start shipping. Think of them as a team that understands the product better than we do.”
“I think different members of our team leverage AI to different degrees. I think it’s a matter of personal preference. For me, I still have trust issues and I need to see the code and I need to know exactly what’s going on.”
What counts as ‘safe to delegate’ isn’t fixed even within a company. It shifts with the task, the stakes, and the operator’s confidence in what the agent is doing under the hood.
“We keep a shared Claude.md—everything we’re working on, plus an architecture file. Before work, we know exactly what to do. It’s like having specialized engineers we continuously train and update.”
“Create a Claude.md with instructions sent on every request—design patterns, anti‑patterns, company context. When the AI makes a mistake, add it so it doesn’t happen again. It’s like onboarding an intern.”
“Our agents use internal endpoints to scrape docs and onboard new APIs. What used to take a day—or a couple of days—now takes an hour or even minutes. We’ve already converted that into a reusable skill.”