Six months of LLM-assisted coding: what stuck and what did not
Notes from using AI coding tools daily on side projects: where they save hours, where they cost them, and the habits that matter.
I have used AI coding tools nearly every day for six months, across several side projects and a few throwaway prototypes. Here is what I would tell my past self.
Where it saves real time
- Scaffolding and glue code. Config files, routing, boilerplate. Hours become minutes.
- Unfamiliar APIs. Getting a first working example is dramatically faster than reading docs cold.
- Refactors with tests. If tests exist, the loop of change, run, fix is fast and safe.
Where it costs time
- Plausible but wrong versions. Models confidently use APIs that were renamed. I now check versions against official docs before trusting any snippet.
- Large unreviewed diffs. A 600-line change I did not read is a debt I will pay later.
- Vague prompts. The output is only as clear as the intent. Writing the intent down is most of the work.
Habits that stuck
- Write the test or the acceptance check first.
- Keep changes small enough to review in five minutes.
- Read every line that touches auth, money, or data deletion.
- Keep a
NOTES.mdof decisions so the next session starts with context.
The product-manager angle
This is the same discipline I ask of AI features at work: define good, measure it, review the output. Using these tools myself made me a more empathetic PM for the people who use ours.