AI-assisted development, in practice
Tools like Cursor and Claude change the shape of the work more than the amount of it. Here's what actually shifted for me day to day.
A year into using AI-assisted tools as a default part of my workflow, the honest answer to "how much faster are you" is: it depends entirely on what kind of work you're doing, and pretending otherwise undersells both the tools and the job.
Where it earns its keep
- Unfamiliar territory. Standing up a new integration, reading through an unfamiliar API's types, or getting oriented in a part of the codebase I don't own — an AI assistant collapses what used to be an hour of doc-reading into a few minutes of targeted questions.
- Mechanical translation. Turning a Figma spec into component scaffolding, migrating a pattern across a dozen similar files, writing the tenth nearly-identical test case — this is where letting a model do the typing is a clear win.
- A second pair of eyes. Asking "what would break this" before opening a PR catches a meaningful slice of edge cases I'd otherwise find in QA or, worse, production.
Where it doesn't
The parts of the job that were always the actual job — deciding what to build, understanding why a system is shaped the way it is, judging whether a shortcut today is a liability in six months — haven't gotten meaningfully faster. If anything, they've gotten more important, because generating code has gotten cheap and judgment about which code to generate hasn't.
The failure mode worth watching for
The easiest mistake is accepting a plausible-looking diff without tracing through why it works. AI-generated code is confident by construction — it reads the same whether it's correct or subtly wrong. The habit that's saved me the most: treat every suggestion as a PR from a fast, capable, context-poor teammate. Review it exactly that way, especially around auth, data mutations, and anything touching production config.
What I'd tell someone starting out with these tools
Use them to move faster through the parts of the job you've already done a hundred times, so you have more attention left for the parts you haven't. Don't use them as a substitute for building the mental model of your own system — that's still the thing that makes you useful when something breaks at 2am and there's no clean prompt for "why is this happening."