Since we started Numyl, the single biggest change in how we work is how software gets built. We build with Claude — Anthropic’s AI assistant — and specifically with Claude Code, the version that sits in a developer’s terminal and works alongside them.
If you don’t write software, the easiest way to think about Claude Code is as a very careful collaborator. It reads the existing codebase, understands what you’re trying to do, writes new code, runs the tests, and shows you what it did. You stay in the driver’s seat — every change is reviewable, every action is explicit — but you cover ground faster than any one person working alone ever could.
Four things change when you work this way.
It’s quicker
Work that used to take half a day often takes twenty minutes. A good example: the site you’re reading this on went from a one-file “coming soon” page to a full marketing site — blog, legal pages, analytics, CI/CD pipeline, SEO assets — in about a week of scattered evening work. A year ago that would have been a multi-week project with real trade-offs about which corners to cut. Now the corners don’t need cutting. We ship the better version because “better” is cheap.
It’s cheaper
The direct cost is modest — Claude costs less per month than most software subscriptions a small company already pays for. The indirect savings are larger. We haven’t had to hire a contractor for most of what we’ve built so far, because the team we have goes further. That lets us spend the budget where it actually matters: on the people who know our business and our customers.
The quality is higher
This is the surprising one. You might expect AI-generated code to be sloppy or shallow. We’ve found the opposite — if you set clear expectations and keep the pieces small, Claude writes code that’s often better than what a hurried human would have shipped. It reads the whole file before making changes. It notices inconsistencies. It writes tests, helpful error messages, and documentation — the things that tired engineers quietly skip. And because Claude can do in minutes what might take an hour, there’s less pressure to cut corners under deadline.
It’s safer
Modern software is hard to secure. A small web application pulls in thousands of lines of other people’s code, talks to dozens of outside services, and has to handle input coming from anywhere on the internet. Most breaches aren’t clever attacks — they’re ordinary mistakes: a missing check, an unescaped input, a leaked credential.
Claude reads code with security in mind by default. When we build a form that takes user input, it reminds us to sanitize and validate. When we touch authentication, it flags permissive defaults and missing edge cases. When we add a new dependency, it notices if we’ve accidentally opened a door we didn’t mean to. Every change gets the kind of careful security pass that only the best code reviews used to produce — and now every change gets it, not just the ones we had time for.
That doesn’t make software unhackable. But it raises the floor: fewer silly mistakes, fewer forgotten checks, and far less code written in a hurry that we’ll regret later.
What hasn’t changed
None of this removes the need for good judgment. Claude doesn’t know what we’re trying to build for our customers — we do. It doesn’t know which trade-offs matter for our business — we do. The job of the person at the keyboard is still the hardest part: knowing what to ask for, and recognizing a good answer when you see one.
What has changed is that the execution — the typing, the plumbing, the careful wiring — is no longer a bottleneck.
That’s why, for us, building with Claude isn’t a novelty or an experiment. It’s how we work. And it’s a big part of why we can build Numyl the way we’re building it: a small team, moving carefully, shipping the full version instead of the minimum viable one.