AI lab
Where I learn to build with AI. A private vault for my notes, lessons, prompts and skills, and a public lab for the parts I choose to share.
A place for small AI experiments: agents, prompts and automations that save real time. It has two halves. The vault is private and is where I work. The lab is public and shows what I publish from it. Both run on Cloudflare next to this site.
The AI vault
The vault is one private store for notes, lessons, prompts, skills and saved links. I read and write it in the browser, also on my phone. Nothing in it is in the repo of this site, and everything in it is private until I publish one item myself.
Search. Every item is searched by its words and by its meaning at once, together with the public pages of this site.
A server for my agents. My Claude Code sessions can search the vault, read an item or a skill, log a lesson, add a TIL, draft a note and save a link. They can never publish, edit or delete.
Sessions become lessons. A script on my Mac condenses a coding session and strips secrets, emails and phone numbers before anything leaves the machine. A workflow then writes a draft: what was built, what went wrong and what I would do differently. I review every draft.
TIL and cards. Short entries I write from the phone, or an agent files for me. Flashcards are made from my own notes, with a short review page for each day.
Two agents. A research agent reads one page I point it at and files a summary as a draft. A review agent reads one of my notes and says what is unclear, what is missing and what is private. Both live inside the vault.
Read later and a weekly digest. A link I save is fetched, summarised and tagged. Once a week a draft sums up what I read, built and learned.
The AI lab
The Lab is the public half. It holds only what I publish, one item at a time. Agents can write drafts, but only I can publish.
Skills. The Claude skills I have written, each with one line to install it, are on the Skills page.
Patterns. The ways of working with agents that I used to build all of this are drawn and written up on the Patterns page.
Evals. The prompts the vault runs are files with test cases. A runner scores them on several models, and the scores are on the Evals page.
TIL. The short entries I publish are on the TIL page.
Notes and lessons. Published notes and lessons from the vault appear in the Lab itself.
How it was built
From one written spec, in nine packages, by agents working in order. Each package had to pass its tests, a run against a private preview and a real browser check before it was committed. The plan, then build, then QA pattern describes how that went.
Where it stands
It is built. It is new, so the public pages are still thin.