We use Claude for coding as a fast collaborator, not an autopilot. It helps us explore a codebase, draft routine code, write tests and refactor, and a developer reviews and tests every change before it ships. We also build products on Claude, such as the Moonlight trip planner.
This is a description of our own practice, not a ranking of tools.
Where it helps us
- Understanding code quickly. Asking it to explain how a module works or where a behaviour comes from.
- Routine code. Forms, validation, API endpoints and data mapping, once we have described the rules.
- Refactoring. Renaming, restructuring and cleaning up with tests as a safety net.
- Tests and documentation. Drafting cases we then check.
- Debugging. Suggesting likely causes from logs and code.
- Content and data work around a site, such as drafting structured pages we then edit.
Where we do not rely on it
- Product and architecture decisions. Those need business context and trade-offs.
- Security-sensitive logic. Authentication, payments and permissions get extra human review.
- Anything involving secrets. Keys and customer data are kept out of prompts. See never paste secrets into an AI assistant.
- Unreviewable big-bang changes. We prefer small, steady steps.
How we work with it
- Give it context, such as the relevant files, conventions and the rules in plain language.
- Ask for small changes, one at a time.
- Read every diff.
- Run tests and the type checker, and check the result in the browser or app.
- Verify claims. If it says something works, we confirm it ran.
- Keep a human owner for every change.
See the full checklist in reviewing AI-written code.
What we have built on Claude
Moonlight, Trripah's AI trip planner, runs on Claude Haiku 4.5 with a custom persona, destination knowledge loaded on demand and lead capture. Read how we built it. We also build document parsing and assistants for other workflows under our AI and intelligent automation service.