My AI-assisted development workflow without cutting corners

AI coding tools make it easier to write more code. That is not the same as shipping better software.

The workflow that has held up for me is simple: use AI to compress mechanical work, but keep architecture, taste, review, and product judgment human-owned.

Where AI helps

I reach for AI tools when the task has a clear boundary:

• exploring unfamiliar APIs
• drafting tests
• scaffolding repetitive code
• comparing implementation options
• refactoring a known pattern
• summarizing large files or logs

That work benefits from speed because the correctness target is visible.

Where I slow down

I do not delegate the product shape. I still decide the architecture, data model, error states, user flow, and what should not be built.

AI can propose options, but the engineer owns the tradeoff.

Tests become the contract

When using AI heavily, tests are not ceremony. They are the contract that keeps fast changes from turning into vague confidence.

For projects like Gorgi, backend tests against real Workers bindings mattered because the system depended on D1, quota checks, and route behavior that mocks would hide.

Small diffs win

The best AI-assisted work happens in small, reviewable steps. Ask for one unit of change. Inspect it. Run checks. Then continue.

Large generated diffs are where quality disappears.

The rule

AI should make good engineering habits faster, not replace them. If the workflow removes review, architecture, tests, or product thinking, it is borrowing speed from the future.

I also speak about this workflow for engineering teams and student groups. See my speaking topics.