My AI-Assisted Dev Workflow: Faster Without Cutting Corners

I have been running an AI-assisted development workflow seriously for about two years. Not as a novelty, but as a professional practice. Here is what works, what does not, and the principles I work by.

The Core Principle: AI as Pair Programmer, Not Author

The most important mental shift: AI does not write my code. It writes drafts that I review, verify, and own. That difference is not just wording. It changes how you use the tools.

When AI is the “author,” you merge outputs without understanding them. When AI is the “pair programmer,” you treat its suggestions the same way you would treat a suggestion from a colleague: worth considering, worth pushing back on, always your responsibility.

This means every AI-generated line goes through the same review I would give any code: does it handle edge cases? Does it introduce a security surface? Does it match the existing patterns in this codebase?

Tools I Actually Use

Kiro for agentic workflows: scaffolding, multi-file changes, and tasks that need context across the whole codebase. It respects project conventions when you set them up properly (AGENTS.md, docs, tests).

GitHub Copilot for inline autocomplete. Fast, quiet, and surprisingly good at picking up project-specific patterns after a few hundred completions.

Cursor for exploratory sessions. I use it when I want to talk through an approach before writing it, or when I am in an unfamiliar part of a codebase and want a thinking partner.

What I Use AI For

Boilerplate and scaffolding. Generating a new API handler that follows the existing pattern, creating a test file structure, setting up a new component with the right imports. AI is fast here and the output is verifiable.

First drafts of utility functions. Give it a signature and a docstring, get a draft, verify the edge cases, adjust. Faster than writing from scratch, and the review process forces you to think through the function’s contract.

Documentation and comments. I almost never write JSDoc manually. AI generates the first draft; I verify it is accurate.

Debugging assistance. Describing a bug and getting a list of possible causes. I do not always follow the first suggestion, but the list narrows the search space.

Regular expressions and complex string transformations. I write what I need in plain English, get the regex, verify it against test cases. Nobody should spend 30 minutes writing regex when there is a better option.

What I Do Not Use AI For

Security-sensitive code. Authentication logic, authorization checks, and anything cryptographic, I write by hand and review carefully. AI-generated security code fails in subtle ways that are hard to catch without domain expertise.

Database migrations. Too much implicit context about production data. AI does not know what is in your database, only what your schema says.

Architecture decisions. AI can sketch options, but “what should the overall shape of this system be” requires understanding of the team, the constraints, the operational environment. That understanding does not fit in a context window.

Final decisions about naming. AI names are often generic. Good naming is one of the hardest things in software, and it is worth the time.

The Productivity Reality

I am clearly faster with AI than without it. Not 10x, though. That number is marketing. It is more like 1.4x to 1.8x on tasks with a lot of predictable structure: tests, boilerplate, docs, and refactoring to a known pattern.

On open-ended design problems, AI assistance is closer to neutral. The value there is more cognitive: articulating a problem clearly enough to ask AI about it often clarifies your own thinking.

The Quality Check

My rule: if AI wrote it, I test it more carefully than if I wrote it myself. Not because AI is worse (sometimes it is better), but because I understand the assumptions in my own code more deeply. Extra tests are cheap insurance.

Before merging any AI-generated code: type-check passes, tests pass, I can explain what every function does. If I cannot explain it, it does not ship.

The Honest Assessment

AI-assisted development is a real skill, not a shortcut. The engineers who use it well are faster and more effective. The engineers who use it poorly ship bugs faster. The difference is in the review discipline.

If you are not using AI tools yet: start. The tooling has crossed a threshold where the productivity gains are real. If you are using them without review discipline: slow down. The cost of low-quality output compounds.