What AI tools we use, and why
The actual toolchain, named, with what each one is for and where each one is not allowed.
By Saurav Kumar Nanda
Most agencies treat their toolchain as a trade secret. We treat ours as a spec. Here is exactly what we use and, just as important, where we do not let it near.
Claude Code, the primary build tool
Claude Code does the first draft of most code: scaffolding, handlers, types, tests for the obvious cases. We run it on the model "claude-sonnet-4-5". An engineer drives it and reviews every line. It is fast hands, not a decision-maker.
Cursor, for review and refactor
Cursor is where we do the second pass: tightening, refactoring, hunting the edge case. Different tool, different mode of attention. The first draft and the review should not be the same motion.
Linear, GitHub, Vercel, Sentry, Plausible
Linear holds the sprint, one card per day. GitHub is the source of truth and the audit trail. Vercel gives you a preview URL for every pull request. Sentry watches production after handoff. Plausible measures whether the thing worked. None of these are exotic. That is the point.
Why publish this
Because a founder paying for engineering deserves to know what is doing the engineering. The agencies that hide the toolchain do it so they can keep billing "senior hours" for tool-speed work. We would rather you see the whole stack and judge the result.
Keep reading
Get the next one.
One argument about shipping software, once a month. No listicles, no more than one email.