
DeskCrew
The agent-native helpdesk for humans + AI agents
I'm building DeskCrew — a support helpdesk where an AI answers from your knowledge base — solo, bootstrapped, no funding, no shareholders. Yesterday I shipped a guided onboarding wizard I was proud of. Then I watched a real person try to use it.
Failure #1: The signup code that was "wrong." They entered their email, got the 6-digit code, typed it in: "That code is wrong or expired." It wasn't. A schema change hadn't actually been applied to production (I took "sql run" as confirmation without checking the verification output), so the server crashed before comparing the code — and my form rendered every server error as "wrong code." The user blamed themselves. The bug blamed them too.
Lesson: error messages that guess are worse than error messages that admit. Now a server fault says "something broke on our side — your code may still be valid," and every failure path logs the real error.
Failure #2: The wizard whose big moment could not physically happen. The wizard's climax: write a help article, ask your own AI a question, watch it answer. Except (a) the quick-create saved articles as drafts and the AI only reads published ones, and (b) the AI feature flag defaults to off for new accounts. Two independent kill switches. The magic moment was impossible for 100% of new signups.
My tests were green the whole time — because my test setup helpfully flipped the flag on and mocked the search. I tested the flow I wished I had, not the one new users actually get. New rule: every onboarding test runs against a factory-default account. No seeded flags. No shortcuts.
What I replaced it with (stolen shamelessly from studying Help Scout, Crisp)
No wizard. You land in the real product.
A demo ticket is already in the inbox — replying to it is the tutorial.
A hosted fake storefront runs your real chat widget before you've installed anything — you message yourself, and it lands in your inbox as a real ticket. That loop is the aha.
A 5-step checklist that checks itself off from real signals (a ticket exists, the widget actually loaded on your domain, the flag actually flipped). No "I did it" buttons. Software shouldn't take your word for things it can verify.
Six weeks of features taught me less than 24 hours of watching one signup fail.
If you're pre-launch: go create a fresh account on your own product today, with a clean browser and none of your admin flags. I'd bet money something in your first five minutes is broken.
(DeskCrew is at deskcrew.io if you're curious what the rebuilt flow feels like — free tier, no card.)
About
Building DeskCrew, the thing I keep coming back to is the trust line. Most "AI support" either over-promises (auto-sends wrong answers) or under-delivers (a glorified FAQ).

1 Comment
The biggest lesson here isn't the bugs—it's that one real user invalidated assumptions an entire test suite didn't. I'd keep validating every "aha moment" against a factory-default account. The faster you prove new users can reach value without hidden assumptions, the stronger the product becomes.