I gave three AIs my LinkedIn archive. They built this website.

Three AIs, three jobs, one set of files

The whole project was a relay. No AI worked from memory of a previous chat; each one picked up files the last one left behind.

1. ArchiveMy 104 posts, 18 project write-ups and my public GitHub repositories. This is the raw material and the source of truth.
2. Evidence map (GPT 6.0)Every post numbered and grouped into 34 themes, with gaps and duplicates flagged, before a single page existed.
3. Content and build (GPT 6.0)Lessons, practice sheets, comics and illustrations written as content files, compiled into plain HTML by one Python script, with link and anchor checks.
4. Editorial pass (Claude)Reviewed the live site against one purpose, merged duplicate ideas, rebuilt the field notes into a learning path and redesigned the reading experience.
5. Hosting (ChatGPT)Configured GoDaddy, the domains and an automatic GitHub deploy, with my passwords kept out of the conversation.
6. Story (Claude Code + GPT Live)A 2:35 instructional video, voiced by two AI actors and rendered on my own computer.

The handoffs were ordinary files: a coverage map, content files, a build script, a validation script, a redirect list, a deployment config and a README describing how to update everything. That is what made the project safe to split between different AIs. When something was wrong I could point at the file, not at a feeling.

GPT 6.0 read the archive before it wrote anything

The first request was not “make me a website.” It was: read everything, then show me what is there. GPT 6.0 had access to my LinkedIn profile and my exported posts. Its first deliverable was an evidence map:

• All 104 posts (August 8, 2024 to July 16, 2026) grouped into 34 teaching themes, with post numbers listed under each theme.
• A check of my recent LinkedIn activity, which found four newer posts the export didn’t include yet.
• An inventory of 18 project write-ups, cross-checked against GitHub so private projects were never presented as public, and placeholder ideas were never presented as finished products.

Only then did it decide what to publish: one practical lesson per theme, a lesson for each newer post, a project lab for each project, and setup lessons for beginners. Each guide got an outcome, ordered steps, a copyable prompt, a way to check your result and a troubleshooting note. It also produced the comics and the illustrated Patrick and Su artwork that runs through the site.

Claude edited for the reader, not for the writer

A first draft built from 104 posts has a predictable problem: it sounds like 104 posts. Several articles taught the same lesson in different costumes, and two guides duplicated each other.

My first prompt to Claude asked it to review the site and “recommend any changes and organization you would make to turn it into an award-winning blog.” The review was good, and it was about looks. My second prompt changed everything: “Re-review the site. The purpose of this site is for people to quickly get the information they need.” With a purpose, the review became a checklist of reader problems instead of a list of design ideas. The Review against a purpose prompt below is that second ask, ready to reuse.

• Every field note now teaches exactly one idea, placed on a numbered learning path in five stages.
• Overlapping articles were merged, and every retired address permanently redirects to the article that now owns the idea, so no old link breaks.
• Three missing foundation articles were written so beginners start with how AI answers, what context it needs and how to frame the problem.
• Duplicate guides were merged (59 became 57), and every article ends with the guides that let you practice it.
• A reading-focused redesign: a one-minute summary, a “Try it” checklist, copy buttons for prompts and a clear next step.

One more edit was pure honesty. The site’s old logo was an orange asterisk that looked suspiciously like Claude’s own icon. The new logo is Su, the sugar glider in sunglasses.

ChatGPT took the GoDaddy work off my plate, inside guardrails

Hosting is where most people stall: FTP accounts, document roots, domains, certificates, WordPress leftovers. I asked ChatGPT to do all of it, and to write down everything it changed.

• A deploy-only account that can reach one folder and nothing else. The old WordPress sites and databases were left untouched.
• Encrypted uploads (explicit FTPS) with certificate checks.
• No passwords in chat. I typed the password into a secure prompt; a setup script tested it, then stored it only as a GitHub Actions secret.
• Every push deploys itself. GitHub Actions validates links and files first, then uploads only what changed.
• Verified, reversible uploads. Each file goes up under a temporary name, is downloaded again and compared by SHA-256, then renamed into place. Overwritten files are backed up. Nothing on the server is ever deleted automatically.
• One canonical address. My older domains now redirect permanently to patrickz.ai.

edit content  →  python scripts/build_learning.py
             →  python scripts/validate.py
git push main →  GitHub Actions: validate, then publish
             →  upload to temp name → verify SHA-256 → rename
             →  live on patrickz.ai

Today, updating the site means editing a content file and pushing one commit. That update path is the real product; the first launch was just its first run.