← Back to journal

Week one — three agents, and the things that almost broke

· In progress · meta, launch, lessons

Two years ago I wrote a book called 50 A.I. Business Opportunities for People Over 50. Somewhere between writing it and last month, I realized I hadn’t built a single one of the opportunities myself. That’s what started this site.

Week one is over. Here’s what shipped — and the things that almost broke.

What went up

Three working AI agents, all free, all honest about what they can and can’t do:

  • AI Side Hustle Match — four questions in, three matched opportunities out, each with a concrete 14-day plan to get your first dollar.
  • Niche Validator — describe an idea in one line. Get an honest verdict, three cheap tests you can run this week, and the three things most likely to kill the idea.
  • First Client Outreach Script — three versions of a first-contact email in different voices, a three-step follow-up sequence, and honest responses to the objections you’ll actually hear.

Plus the book funnel at /companion, a three-tier consulting page at /consulting, and this journal you’re reading.

What I got wrong

I want to catalog the mistakes because this is a lab, not a launch announcement. If I only wrote about what worked, I’d be doing the exact thing I built this site to avoid.

I picked the wrong model the first time. The original deploy used Claude Sonnet for the matching agent. Sonnet is genuinely capable but reliably takes 12 to 15 seconds per response. Netlify’s free-tier function cap is 10 seconds. Every agent run returned a 504 timeout to the user. Switched to Claude Haiku 4.5, cut response times to about 8 seconds, and quality turned out to be more than enough for “pick three from fifty and write short plans.” The lesson: the biggest, most expensive model is rarely the right one. Match the tool to the job.

I over-engineered the auth. I spent most of day one building magic-link sign-in on Supabase, a paid subscription tier on Stripe, rate limiting per user, saved run history. Then I looked at what this site is actually for — a consulting lead funnel with a book as the top of the funnel — and realized I had built the skeleton of a SaaS, not a funnel. I ripped out three services (Supabase, Stripe, and a transactional email provider) in a single commit. The site now uses Kit, which I already use for another business, and no user accounts at all. Sixty percent less code, one hundred percent more aligned with what I’m actually trying to do.

I almost nuked working DNS. After pointing the domain at Netlify, I saw IP addresses in the lookup I didn’t recognize. They looked like AWS addresses. I nearly deleted every record and started over. A sanity check stopped me: Netlify runs on AWS. Those addresses were its load balancer. The site had been serving correctly the entire time. The lesson I already knew but had to relearn: before deleting infrastructure, check whether it’s actually broken. Most of the time it isn’t.

I assumed too much about Kit. I tested the /companion signup with my own email and never received the welcome. I spent an hour debugging what I thought was a broken integration before checking a mundane fact: my Gmail wasn’t syncing on my phone. The integration was fine. I was looking at the wrong inbox.

I missed the obvious feature for a week. The agents were wide open — anyone could use them without giving an email. I built the whole architecture so that every agent run would also be a lead, then forgot to gate it. I caught this myself eventually. Now every agent run captures an email before the tool runs, and those subscribers become part of the book list. It’s a small change and it probably doubles the lead capture of this site.

What I’m watching over the next two weeks

  • First real subscribers through /companion — will anyone actually trade an email for the PDF?
  • First agent runs from strangers — does the output land, or does it need more calibration?
  • First consulting inquiry — the honest signal that the funnel is working end-to-end.

I’ll post what the numbers actually say. If zero humans show up, I’ll say that too. That’s the point of the lab.

The next agent ships in two weeks. Probably a Newsletter Starter Kit — maps to opportunity number two in the book. I’ll write a post about whether it actually helps anyone start a newsletter that gets read.

— Jason