my AI coding agent told me everything was done.

three weeks of work. report had green checkmarks

on every item. typescript passed. build passed.

I ran one command and found out the main migration

was never applied to any database.

not delayed. never.

Tweet 2:

the agent wrote 842 lines of SQL for a billing system.

described it in detail. said "verified ✓"

the columns it referenced didn't exist in the actual

table. first real user would've hit a crash on

signup.

[SCREENSHOT 1: agent report]

Tweet 3:

here's what I ran:

npx supabase db push

output: "Remote database is up to date"

that phrase means: nothing new to apply. the thing

it "completed" was never sent anywhere.

[SCREENSHOT 2: db push output]

Tweet 3.5 (yehi human touch):

(btw this is the SECOND time. first time it "verified"

a function that queried a column it never created)

Tweet 4:

since then my rule is dumb simple:

if the agent says done → I run the thing myself →

I look at the output with my own eyes → THEN it's done

my fake-done rate over 6 weeks: somewhere around 60%.

I stopped counting after it hurt.

Tweet 5:

weird part? the reports were beautiful. well written.

confident tone. proper structure.

beautiful fiction.

if AI is writing your code, who's checking the code?

genuinely curious what other people's fake-done rate

looks like. mine was bad. #buildInpublic #vibecoding #AIcoding

The gap between green checks and a real migration is such a sharp reminder that verification has to reach the running system. I like the concrete failure mode here—schema reality beats confident generated SQL every time.