Building · Nov 2026 · 3 min

Failures must be loud

A user told me they couldn’t sign in to Faddy, my form builder.

Faddy signs you in with a code sent by email. The code never came. I checked, and my email provider, PostShiba, had been refusing every single email for 10 days because it saw my sending domain as unverified. The reason was one line in the logs: domain_unverified.

Ten days. No alert, no error on screen, nothing in my inbox. Nobody knew until a real person tried to log in and told me.

Then I found a second one

While I was in there, I found another silent failure on the same screen.

The sign-in page runs Cloudflare Turnstile, a bot check. The page’s security policy was blocking it. So the “Send code” button just stayed grey. No message, no hint. You’d think the site was broken, and you’d be right.

Two outages, both on the most important screen in the product, and both completely quiet.

The rule I wrote that night

I added one rule to Faddy’s project rules, the file every change has to follow:

Failures must be loud.

What that means in practice:

  • No silent fallbacks. When something a user needs fails, the screen says what failed and offers a retry. A greyed-out button is never the only sign.
  • Every flow reports success and failure. Sign-in, sign-up, invites, publishing, submissions and every email send an event when it works and another, with the reason, when it doesn’t.
  • A failure that blocks users alerts me at once. On Discord and by email, within seconds. A log line nobody reads doesn’t count.
  • Every new outside service gets an alert and a health check the day it’s added.

What I shipped the same evening

A backup email provider. When PostShiba fails as a whole, the same email goes out through Cloudflare Email Service instead, and I get an alert saying the primary is failing. Users keep getting their codes while I fix the real problem. A bad address never falls back, only a provider failure does.

Instant alerts. Emails failing (with the provider’s reason and the fix), a script blocked by the security policy, the bot check unreachable, refused submissions, bursts of server errors. They go to a Discord channel and to my email, deduplicated so one problem doesn’t become a hundred pings.

PostHog for errors and the sign-in funnel. Code requested, sent, failed with a reason, verified. By user id only: no emails, names, IPs or form answers.

A sign-in screen that tells the truth. If the security check can’t load, it says “Couldn’t load the security check” with a Try again button.

More ways in. By one in the morning you could sign in with Google, a passkey or a password. Email codes still work, so nobody gets locked out because one service had a bad day.

The user comes first here, in two ways. A user’s report goes to the top of my day, ahead of whatever I had planned. And while I fix a provider, users keep getting served. One person’s “I can’t log in” turned into a whole evening of work, and every Faddy user is safer for it.

The stack behind it

For anyone building something similar, these are the services behind those releases.

Sign-in

  • Google: Sign in with Google (OAuth / OpenID Connect)
  • Passkeys (WebAuthn): work with Apple iCloud Keychain, Google Password Manager and other passkey managers
  • Cloudflare Turnstile: the bot check on sign-in

Email

  • PostShiba: primary email provider
  • Cloudflare Email Service: backup email provider

Monitoring and alerts

  • PostHog: error tracking and product analytics
  • Discord: instant alerts through a webhook, plus alert emails through the providers above

Hosting and shipping

  • Cloudflare Pages: the console and public forms
  • GitHub Actions: CI on a self-hosted runner

Testing

  • Playwright: browser end-to-end tests, including a virtual passkey device

And the open-source pieces: webauthn-rs (passkeys in the Rust API), Argon2 (password hashing), Axum (the Rust API) and React (the console).

How I built it

I built all of this with Claude Code, inside Orca, the editor I use every day. I wrote the rule, decided what each alert should say and who it goes to, and tested every path: a fake Google, a software passkey, a provider that refuses everything.

The tools made it fast. The rule is what I’ll keep.

If your product has a screen people can’t do without, ask yourself one thing: if it broke tonight, who would tell you first? If the answer is “a user”, that’s where to start.