We killed Co-work today as the primary harness and moved everything to Claude Code. Co-work was genuinely the right call a week ago. Today it is the wrong one, for two hard reasons.
One, Claude Code now ships native schedule and dispatch. Those are the exact capabilities we were paying the Co-work coordination tax for. There is no reason to pay it twice. Two, Co-work runs in a sandbox, which means git push cannot happen automatically from inside a session. Every handoff was a manual copy and paste back into the real repo. Fine for experiments. Fatal for a daily shipping cadence where the PM agent, the writer agent, and a human are supposed to converge on a single branch.
No migration pain. The repos were already on GitHub. Every MCP we needed was one install away. Everything below this line in today's log was shipped on Claude Code.
The moment the platform catches up with the reason you adopted a tool, the math flips. If you are still defending last month's decision this month, you are paying rent on a room you are not using. This is the same muscle a Chief AI Officer has to build: audit the stack quarterly, not emotionally.
Point 02: Somebody Bought the Seat
Mid-day, a client bought a 5-day Infinite Leverage Blueprint engagement. Five days on site with the team, building their own version of this system, walking out with their agents and their playbooks in their own repo. I want to sit with that for a second, because the timing was comedy and the implication was not.
I am still building the program. Day 4 of 88. And someone picked up the phone and said, I am in. That changes the job for the remaining 84 days in one small but permanent way. We are no longer building this just for us. Every piece of infrastructure we wire up from here on has to be teachable, reproducible, and walkable with a non-technical founder sitting next to us. Code that only I understand is technical debt against a human who is flying here on their own dime.
The best forcing function in a startup is a paying customer who is on their way. Most of what we did today turned out to be exactly the curriculum we will need for that week. The teaching work and the building work collapsed into the same work, which is the only version of this that ever scales.
Point 03: Teaching the Non-Techie Is the Product Now
Most of Day 4 was not on the core project roadmap. Most of Day 4 was on leveling up the non-technical operator (that is me) on the plumbing that every modern web project needs. One hour with Supabase, one hour with contact forms, one hour with clean URLs, and a few smaller passes on copy, social share cards, and image polish.
Halfway through the Supabase setup, Claude handed me instructions using the old names for Supabase keys. Supabase had quietly renamed them. The old name "anon" is now "publishable." The old name "service_role" is now "secret." Claude's training data had not caught up. I only noticed because I opened the Supabase dashboard with my own eyes and the words on the screen did not match the words in the chat. That is the whole lesson. AI is the best junior teammate you have ever had, and you still have to look at the work.
Later, when I asked Claude to design the place my form data would live, it proposed three separate tables. One for general inquiries, one for keynote inquiries, one for consultation inquiries. Correct developer instinct. Wrong strategic call for my life. I told it no. One table called leads, with a column called inquiry_type. One query to count the week, one place to look, one bucket. Claude flipped the design without complaint. It was the right call, and it was a human call.
Trust the speed. Verify the details. Stay the editor. An experienced human architect, even for one sentence of judgment, is the cheapest insurance in the building. This is the exact muscle we teach inside the AI Officer certification, because it is the skill that separates "using AI" from "leading AI."
What We Actually Shipped
For the people who want to see the receipts, here is the compact version. No hero lap, no drama, just the list.
- Moved the whole operation to Claude Code. Killed Co-work as the primary harness.
- Migrated a static site to clean URLs. Added vercel.json with cleanUrls true, stripped .html from 652 internal links across 41 files, and kept the old .html paths alive as 301 redirects so no backlinks break.
- Wired a contact-form backend to Supabase. Server-side client using the new secret key (not the old service_role name), rows confirmed landing in Postgres.
- Designed one leads table with an inquiry_type discriminator (general, keynote, consultation) instead of three tables. One query, one answer, one bucket.
- Built a unified /api/inquiries route. Field whitelist at the door, per-type validation, name and email always required.
- Shipped a reusable InquiryForm component. Schema-driven, so every future form on the site uses the same plumbing.
- Launched a book-keynote page and a schedule-consultation page with hero, pricing strip, social proof, Lark scheduler embed, and full inquiry form.
- Rewired navigation site-wide so hero buttons, navbar, and ContactCTA all point to the new pages.
- Published a new featured blog post on Supabase for non-technical founders, written in the letter voice, live and shareable.
- Fixed social share cards site-wide. metadataBase on the root layout, openGraph and twitter (summary_large_image) on every post.
- Closed the laptop at 4:00 PM to go teach my day-job class. Day job first, always.
What's Next
Day 5 is about finishing what Day 3 said Day 4 would finish. Complete the writer agent. Complete the designer agent. Get the URL mapped. Get the website updated. The comms-layer work from Day 3 is still open, and the Infinite Leverage Blueprint week is on the calendar now, which means the documentation tax is no longer optional. Every agent we ship from here has to come with a one-page explainer a non-technical founder can actually read.
Day 1 was vision. Day 2 was foundation. Day 3 the machine started moving. Day 4 we picked the harness. Day 5 is when something the agents built has to show up on the open web with a name on it.
The Real Insight
The three lessons of Day 4 are one lesson in three outfits. Pick the tool that pays you back, not the tool you picked last month. Let a paying customer rewrite your roadmap for the better. And when you are teaching yourself, you are usually also writing the curriculum for somebody else.
None of these are technical skills. They are the leadership muscles a Chief AI Officer has to build. The job is not prompting better. The job is deciding what to build, what to kill, and what to hand to a human walking in the door next week.
Read next: Day 3: The System Started Running · Day 2: Set Up Trauma · How to Build a 1-Person Company With AI
Why did you abandon Co-work for Claude Code?+
What is the Infinite Leverage Blueprint engagement?+
How can a non-technical founder ship real web infrastructure in one day?+
What is the trust-then-verify rule for AI pair programming?+
Dave Hajdu is the founder of the AI Officer Institute and Edge8 AI. He works with founders and executives across more than 20 countries to build the leadership capabilities the AI era demands. Learn how to build your own AI team at caiocoach.com.