4 min read · Starter draft
Where AI Helps My Workflow, and Where It Doesn't.
A working list, kept honest by the things that went wrong.
I use AI tools every working day, and I have stopped having a general opinion about them. They are very good at some jobs and quietly bad at others, and most of the skill is knowing which job you are holding.
This is my current list. It changes.
Where it earns its place
Exploring options. I can ask for five structurally different approaches to a layout or a data model and have them in minutes. Most are wrong. Seeing why they are wrong is faster than reasoning my way there from nothing.
Work with a clear right answer. Converting a design token file, writing the tenth variant of a component, migrating a set of files to a new pattern. I can verify the result quickly, so the speed is nearly free.
Tests and edge cases. It is good at the question "what inputs would break this?" and patient about writing them all down.
Organising research. Clustering interview notes into candidate themes is a real time saver, provided I treat the clusters as a first draft and go back to the transcripts.
Explaining unfamiliar code. A fast route into a codebase I did not write.
Where I keep my hands on it
Deciding what to build. It will happily help me build the wrong thing very well.
Architecture. Its suggestions are locally sensible and globally unaware. It does not know what the system will need in six months, and it rarely says so.
Interaction detail. Timing, easing, what happens when you interrupt an animation halfway. This is feel, and I have to feel it.
Anything I cannot verify. If I could not tell whether the output is correct, I should not be delegating it yet. I should be learning it.
The failure I watch for
The dangerous output is the plausible one. Broken code announces itself. Code that works, reads cleanly and encodes the wrong assumption will pass review and wait.
So my rule is that I own everything that ships, whoever typed it. If I cannot explain a line, it is not ready.
The net effect
I am not doing less thinking. I am doing the same thinking across more attempts, with the tedious parts removed. That turns out to be a large improvement, and a different one from what the hype describes.