How We Use Claude to Build Entire Codebases Autonomously

Over the past year, we've integrated Claude into our development workflow in ways that go beyond code autocomplete. We're using it to scaffold entire features, write migrations, generate test suites, and produce documentation — autonomously, with human review at each checkpoint. Here's what works, what doesn't, and what it means for how we build software.

What We Mean by Autonomous Codebase Work

Autocomplete (GitHub Copilot, Cursor) suggests the next line. That's useful but incremental. Autonomous coding means giving Claude a spec — "build a multi-tenant subscription billing system with M-Pesa and Stripe, with these data models and these business rules" — and having it produce the full implementation: migrations, models, controllers, Livewire components, tests, and documentation.

We're not running Claude unsupervised in production. Every output gets reviewed by a senior engineer before it ships. But the time from spec to reviewable code has dropped by 60–70% on well-defined features.

Where It Works Best

Claude excels at CRUD-heavy features with clear patterns. Filament admin resources, API endpoints with standard REST patterns, database migrations, form validation logic, and repetitive Blade templates — these are where Claude's output quality is highest and review time is lowest.

It's also exceptional at writing tests. Given a controller or service class, Claude produces comprehensive feature tests that cover happy paths, edge cases, and failure modes. Writing tests is the task developers most consistently skip under deadline pressure — having Claude generate a solid test scaffold that a developer reviews and fills gaps in changes the quality baseline.

Where Human Judgment Is Still Essential

Architecture decisions, security-sensitive code, and anything involving money need careful human review. Claude will produce working M-Pesa STK Push code, but it won't catch the Safaricom-specific edge cases we've learned from production — double callbacks, MSISDN hashing in C2B, the unreliable sandbox. That institutional knowledge lives with the engineers, not the model.

Performance optimisation also requires human judgment. Claude produces correct code but doesn't always produce the most efficient query structure for complex relationships. N+1 query patterns in Eloquent still require an experienced eye to catch.

The Workflow We Use

For a new feature, we write a detailed specification: data models, business rules, user flows, edge cases, and integration points. We feed this to Claude with relevant existing code as context. Claude produces an initial implementation. A senior engineer reviews, revises, and commits the result.

The spec-writing step is the one most teams skip when adopting AI coding — and it's the one that determines output quality. Vague input produces vague output. A well-structured spec with clear acceptance criteria produces code that requires minimal revision.

Impact on Our Team

We're not smaller because of AI tooling — we're more ambitious. Features that would have taken two weeks now take three days. That means we can take on more complex projects, build more thorough test coverage, and invest time in the things that benefit most from human expertise: client relationships, architecture decisions, and the Kenya-specific knowledge that no model has.

The developers who thrive in this environment are the ones who are good at specification writing and code review — the senior skills. The ones who struggle are those who relied heavily on writing code from scratch. The job is changing: the premium is on judgment, not keystrokes.