How I run this site
This site is built and run by Claude, and powered by my voice.
This site is my working example of how I think AI-enabled teams should work: a team of AI agents writes, builds and publishes it, guardrails keep the quality right, and nothing goes live until I have approved it.
Powered by my voice
Almost everything here starts with me talking. I dictate with superwhisper, so I can hand over all the context I have, quickly: what I want, why it matters, and what I have seen that should shape it. I guide the site by voice, and Claude reads my notes for intent rather than word for word, because dictation is rarely tidy.
Letting go of writing the code myself was a real shift. What I do now is set the direction, give the context and hold the work to account, which is much the same shift I think engineering leaders are making with their teams.
How a change flows
Almost every change, from a new post to a new page, takes the same route.
Steps marked Me are mine; the rest are agents.
- MeA voice noteI talk it through: what I want, why, and anything I have seen that should shape it.
- The conductorReads it for intent, writes a spec card, and picks the agents and checks the change needs.
- The specialistsStudio, build or voice agents do the work, each with one job and one guardrail.
- The gatesOne command runs every automated check. Nothing moves on while any of them fails.
- The criticsA reviewer, design critic or narration critic, starting fresh, tries to break it.Anything they find goes back to the specialists, and round again.
- A previewEvery change gets its own preview link, away from the live site.
- MeMy reviewI look at it myself, every time, and say “ship”. Nothing merges on its own.
- Live
An orchestration of agents
Behind the site is an agentic workflow. A conductor reads each request and decides which specialists it needs. The specialists cover design, building, writing, accessibility, speed, security and search, and a narrator looks after the read-aloud posts. Each has one job and one guardrail.
Three independent critics, for code, design and narration, start fresh and check the work before I see it. They never check their own work.
The gates
One command runs every automated check before anything can ship: types, linting, unit tests, a production build, performance budgets, and browser tests that include accessibility checks in light and dark and a check that every old address on the site still works.
Automated checks catch a lot, not everything. So I always review every change myself, on its preview, before it goes anywhere. Nothing merges automatically.
The guardrails
- No change goes straight to the live site. Every change is reviewed as a pull request, with its own preview link to view it, separate from the live site.
- The agent that made something never approves it. Code, designs and narrations each have their own independent critic.
- Facts about my career go live only once I have confirmed them: the live build refuses to publish while any are still waiting for me. Nothing is invented.
- Paid generation, such as a narration, needs my approval of a written estimate first.
- Accessibility (WCAG 2.2 AA, checked in light and dark) and the performance budgets block shipping.
Images and voices
The 3D objects on the home page, the gimbal, the rig and the notepad, are Blender renders made for this site, played back frame by frame as you scroll.
Some posts can be read aloud. The narration is an AI voice made with ElevenLabs, directed by the narrator agent and checked word for word against the post by the narration critic. It is an experiment and a proof of concept, not a promise for every post.
Want to see what it produces? Read the posts, or follow the AI journey.