At a glance
WHAT IT’S REALLY ABOUT
Gusto’s CTO ships Cofounder in 10 weeks using Claude Code
- Gusto Cofounder was built by four engineers and one designer in 10 weeks, going from a layover-built prototype to a tier-one launch inside Gusto.
- The team intentionally removed traditional process artifacts—no Jira, no specs, no Figmas, no standups—replacing them with a 24/7 perma-Zoom and extremely fast code review cycles.
- Product scoping happened through “build-and-decide” workflows where finished pull requests functioned as proposals, and unwanted features were routinely deleted because code generation lowered the cost of iteration.
- The technical architecture was deliberately simple: a Cloudflare Worker running the agent loop with Vercel AI SDK, plus in-house tools and straightforward “memory” stored as database writes.
- Cofounder’s user value centers on multi-channel interactions (web, SMS, Slack) and real operational workflows like payroll preparation by pulling data from systems like Google Sheets and computing business-specific rules before submitting payroll.
IDEAS WORTH REMEMBERING
5 ideasCutting process can be the process for zero-to-one builds.
The team moved fastest by removing meetings, specs, Jira, and even Figmas, relying on a single shared whiteboard and constant collaboration in a perma-Zoom to keep decisions synchronous and lightweight.
Use pull requests as the new PRD and scope gate.
Instead of debating hypothetical requirements, they built real, review-ready PRs and then decided whether a feature deserved to exist; if not, they closed or deleted the work with minimal sunk-cost pain.
Ship “fake” UI to production early, then breathe life into it.
The designer shipped a functional-looking but canned-response experience behind a flag, while engineers built the agent loop and data models in parallel, gradually replacing mocked behavior with real functionality without rewriting the UI.
A simple agent loop stack is often sufficient.
Cofounder ran the agent loop in a Cloudflare Worker and used Vercel AI SDK, avoiding heavy frameworks for planning/memory; “memory” was treated as a tool that writes to a database column, keeping complexity low.
Fast PR review is a force multiplier—especially with AI-generated code.
Median PR review time was ~9 minutes on the team, enabled by perma-Zoom and real-time walkthroughs; this prevents AI-assisted coding from creating unreviewed backlog and keeps quality aligned with team taste.
WORDS WORTH SAVING
5 quotesWe had no meetings, we had no tech specs, we had no Figmas, we had no Jira board where we tracked stories or tracked work. Uh, we had no standups, no retros. We had nothing.
— Eddie Kim
It was, like, literally just Zoom, a lot of Claude Code tokens, and some, like, really passionate people about turning this thing in- into a reality.
— Eddie Kim
We would build features, and we'd just have a discussion, like, "Does this make sense to have or not?" If it is, then, like, it would get code reviewed right then and there. And if not, like, we would just delete it.
— Eddie Kim
The cost to write code is now so low that you can actually, uh, build, uh, products in, in this way and whereas I think you couldn't do that, you know, six months ago.
— Eddie Kim
I think for a subset of projects, like kinda more zero to one, I think docs are absolutely dead.
— Eddie Kim
High quality AI-generated summary created from speaker-labeled transcript.
