Does your AI coding assistant also say “Done!” when it's definitely NOT done? 😂
by•
I started building my own digital marketplace with zero coding experience, using AI to guide me through the process.
And I’ve discovered something hilarious (and occasionally terrifying). 😂
AI coding assistants can sound SO confident.
“Everything is working correctly.”
“All issues have been resolved.”
“The implementation is complete.”
Then I open the website, click one button, and… surprise! Something is broken. 😂
Now I use Claude Code and Codex together, partly because I want them to challenge and review each other’s work.
But even then, I’ve learned that two AI agents agreeing doesn’t necessarily mean they’re right.
Sometimes I feel less like a founder and more like a detective investigating what my AI developers have been up to. 😂
I’m curious: What’s the funniest or most frustrating thing an AI coding assistant has confidently told you was fixed… when it absolutely wasn’t?
I have a feeling we’re going to hear some good stories. 👀
14 views
Replies
All the time. So I stopped letting it just say it.
There's a hook that runs every time the agent finishes a reply. If it says something is fixed without a line saying how it checked that under the same conditions I had, the reply gets blocked and it has to rewrite it. If it didn't check, it has to write "unverified". That hook caught it twice tonight.
It also has to look at the real screen itself before asking me to check anything. The project management app I run has about 150 Playwright E2E specs for that.
It declared an analytics task done and generated a clean report. The metric looked completely plausible, but it was hardcoded mock data never wired to the database. It took a full day to realize nobody had measured anything. Inspect the diff for static values first.
Very familiar 😄 What fixed most of it for me: "done" in my project means the agent built the app, installed it on the Android emulator, tapped through the screen and looked at a screenshot. Not "it compiles", not "the tests pass". It screenshots before every tap too, because the debug build lags and it used to tap blind and report success on a screen that hadn't even loaded yet. Funny thing is it can be wrong the other way too. A freshly deployed backend trigger quietly ignores events for its first minute or two, and the agent once spent a while "fixing" code that was fine, the trigger just wasn't awake yet. Now that's written in the notes.