Does building with AI save you time or just move the work to debugging?
by•
I recently finished shipping a project using mostly AI prompts. Getting the basic app running took almost no time which was great but getting it ready for real users was a completely different story.
I ended up spending days catching small mistakes that the agent missed like wrong timezone math missing input rules and broken uploads.
You really cannot trust it when it says a bug is fixed until you test it yourself.
What has been your biggest time sink when building products with AI tools?
7 views
Replies
An agent built a report with a hardcoded metric that looked completely real. Nobody had actually measured it. It took a day of chasing downstream logic to find out the model had just stubbed in a plausible placeholder instead of querying the database.
For me it moved the work, but to a different place. The bugs weren't in the parts the agent struggled with, they were in the parts it was confident about: dates and timezones, retries, anything touching money or a third-party API.
What helped most was boring: write the check before asking for the fix (a test, a query, a curl), and don't accept "fixed" until that check passes on real data, not a fixture. And keep diffs small. A 30-line change I can read in a minute; a 600-line "cleaned up a few things while I was there" is where the days go.
Konstantin's placeholder metric is the scariest version of this. Has anyone found a reliable way to make agents say "I didn't measure this" instead of filling the gap with something plausible?
both honestly. writing got way faster, but now most of my time goes into reading diffs and running tests before anything ships. still a net win for me