Common mistakes
Some requests stall for reasons that have nothing to do with the request being hard.
Sending to Slack, Notion, HubSpot or Salesforce is written but has never been used against the real service. KakaoTalk and SMS still need Korean carrier review and customer consent. Company sign-in is not connected. Plan around that before you promise it to anyone.
Not saying what the business is
Ask for a nice-looking site and you get a nice-looking site. It just will not be yours.
One word about the trade changes everything. A cafe, a language school and a clinic need different screens and a different tone.
Twenty things in one go
Ten screens and twenty features in a single request makes it hard to tell what is missing when it comes back.
Start with the three or four screens that matter. Add the rest while looking at the result.
Waiting on a card nobody pressed
Large jobs show a confirm card before they start. Until you press it, nothing runs.
Our own test harness once walked past that button and reported the build as timed out. The card was sitting there on screen the whole time. If nothing seems to be happening, look for it.
Repeating the same words
Saying it again the same way gets you something similar again. Say what should be different.
"I do not like the booking screen" is weaker than "on the booking screen, let people pick the date from a calendar".
When it stops on its own
If a check fails, it repairs its own work — three attempts, then a stronger model and two more.
After the fifth it stops and tells you what it was stuck on. Half the credits come back, and you can send that job through once more without paying again.
Things we will not build
Investment schemes that promise a return, gambling, and medical advertising without the qualifications to make it. Asking does not help.
It is blocked in the finished code, not only in the instructions. These pages can put the person who published them in front of a prosecutor, so it is not left to chance.