I use AI in the editor every day. I still read the diff. The tool is fast at a first draft and confident when that draft is wrong.
Completion is a suggestion. The commit is mine.
What these tools are
- AI code editors — the whole editing surface can chat with the repo, edit several files, and run commands if you let it
- Code completion tools — inline suggestions as you type. Best for patterns the codebase already has
- Development tools around them — repo context, terminal access, and rules files. The useful ones follow project conventions. The useless ones invent a second architecture in one prompt
How I use them
- Fill in a function when the signature and the tests already say what “done” means
- Explain an unfamiliar API, then check the docs anyway if the cost of being wrong is high
- Draft tests I will actually run
I do not let an agent migrate a database, change auth, or “clean up” a folder I have not named. Scope is the feature. A wide prompt is how you get a wide bug.
What I would study next
This masterclass is the product path: use a model, ground it, constrain it, measure it, ship it.
If you want adjacent depth:
- Prompt engineering, as a craft, once your prompts are still failing for reasons you can point at
- AI and data scientist work — training, classical ML, evaluation at the model layer — if you are moving toward building models, not only products on top of them
I would not collect certificates. I would ship one feature with evals and a cost number you can defend.