Most failures never reach you. Errors here are written for the assistant rather than for a person — they say what to change, not just that something went wrong — so the usual outcome is that your AI reads one, fixes the code and tries again without mentioning it.
The ones that do reach you are the ones it can't fix alone. There aren't many.
"That name is taken"
Someone already has it. Names are global — one bakery across everyone — so
short common words went years ago.
Pick something more specific. sunrise-bakery-manchester is available in a way
bakery never will be. Your assistant should suggest an alternative rather than
quietly appending a number, because bakery-4 is not an address you want on a
business card.
Nothing is broken. Say a different name.
"You need to verify your account"
You asked for something that runs code on a server — saving form submissions, a login, a database, an API. Static pages publish immediately; server-side code needs a verified account first.
Verification is a $0 card check or a text message, and it takes a minute. The card isn't charged; it's a check that a person exists.
Worth repeating because it catches everyone: upgrading doesn't fix this. Paying raises your quotas. Verifying is what allows server-side code. They're separate on purpose — running untrusted code is the abusable thing, so it's gated behind proving you're real rather than behind a payment.
Often the better move is asking whether you need a server at all. A contact form that opens the visitor's email app is static and works instantly.
"You've used your monthly allowance"
The free plan covers 100,000 visits a month. Past that, sites return an error until the month resets.
Two things people get wrong here:
Publishing again doesn't help. The site isn't broken, and new code won't bring it back. It's a quota, not a bug.
You aren't being billed. We stop serving rather than running up a bill you didn't agree to. That's worse for your uptime and better for your bank account, and it's the trade we chose deliberately.
Your dashboard shows the reset date. Or upgrade, which lifts it immediately.
"That package isn't available"
Your AI wrote code importing something from npm. Packages aren't installed here — there's a curated list instead:
hono, itty-router, zod, @neondatabase/serverless, @upstash/redis,
jose, bcryptjs, nanoid, uuid, marked, date-fns, ms.
The error names what's available, so your assistant should rewrite using one of those or write the logic directly. This one it can genuinely fix alone — if it comes back to you saying "the package isn't supported", tell it to use an alternative from the list in the error.
Why the restriction: installing a package runs whatever install scripts it ships with, on the machine that holds the database credentials. That's not a thing to do on demand.
"That isn't supported here"
Something in the code needs a capability this runtime doesn't have. Usually one of:
- Reading or writing files. There's no disk. Store data in a database.
- Sending email directly. Email needs a protocol we deliberately don't allow, because that's how a platform becomes a spam source. Use an email API.
- Long-running background work. Apps respond to requests; they don't sit running between them.
- Raw network connections. HTTP to allowlisted hosts only.
The error names the replacement. Your AI should take the hint and rewrite — that's the case this whole system is built for.
"The site is live but it's throwing an error"
The deploy worked and the code is failing at runtime. Different problem entirely, and a much more ordinary one — it's a bug.
Ask your assistant to check the logs. There's an endpoint for exactly this and the log names the actual exception. Guessing at a fix without reading it wastes everyone's time, including the AI's.
"Your AI says it can't connect"
Almost always the code. Two cases:
It never worked. The code was mistyped or truncated. Get a fresh one and paste the whole line.
It worked and stopped. The code was replaced or deleted — maybe you made a new one and this conversation still has the old.
Either way: get a new code and give it that. What won't work is looking up your existing one. Once a code has been used it isn't shown again, deliberately, so it can't be read off your screen later. If your assistant asks you to "check your codes page and paste it again", it's asking for something that doesn't exist. Get a new one instead.
"Sites are paused" with nothing on your account
Rare, and not you. There's a platform-wide spending limit that stops everything if costs spike unexpectedly — a safety valve so a runaway bill can't happen.
Deploys still work; sites return an error. Nothing you change fixes it. Contact support and it'll be dealt with.
When you're properly stuck
Ask your assistant to check whoami. It reports your plan, what your account is
allowed to do, how much allowance is left and what specifically is blocked — in
a form written for the AI to act on rather than for you to interpret.
It's one call and it usually turns "something's wrong" into a specific sentence. Which is the whole point: an error that tells you what to change is worth more than one that tells you something failed.