The situation
In 2026, a marketing leader who cannot work with AI directly is managing something they do not understand. Having a point of view on AI is not fluency. Approving the budget for it is not fluency. I mean ground level: hands on the tools, building something real, close enough to tell when the output is wrong.
That applies at every level of a marketing organization, which means it has to start with the person running it.
The idea
So I keep my hands on the tools. At Elementum that meant making a two-person team behave like a larger one, internally and under deadline. This site is the same instinct pointed somewhere public and permanent — a real property on a real domain, maintained by me rather than briefed out.
What it took
- The agent. Claude Code, Anthropic's terminal-based coding agent, working directly against the repository: reading files, writing them, running builds, and committing.
- The stack. Next.js 16 and React 19 in TypeScript, Tailwind CSS v4 for styling, MDX for content, Git and GitHub for version control, Vercel for hosting. Every push to the main branch deploys itself.
- Separating content from code. All the prose lives as MDX files in a content directory, apart from the components that render it. Adding a case study means adding a file. That is the difference between a site I can update and a site I have to ask someone else to update.
- Teaching it the project. Two files carry the weight: one holding the design rules, architecture, and conventions, and one holding my career facts so the agent quotes rather than invents.
- Checking the work. Every change is built, type-checked, linted, and the affected pages fetched before it ships.
The first version went from an empty folder to a deployed site in a single afternoon. Everything since has been a handful of working sessions across a week.
The outcome
A site I can change in minutes, and a much sharper sense of where these tools are strong and where they are confidently wrong.
Knowing where they are wrong is the more valuable half. Twice this project produced something confident and false: the agent told me the site was not hosted anywhere, when it had been deploying on every push for days, and it described a job I held in terms I had to correct. Both were plausible. Neither was true. Catching that in your own domain is the skill, and no vendor briefing teaches it.
This is a work in progress and will stay one. Case studies are still being written and copy is still being tightened. I would rather publish honestly mid-build than wait for polish.
What I would tell another CMO
Do not delegate your own fluency. Your team will use these tools whether or not you understand them, and the only question that matters is whether you can tell good work from work that merely sounds right. Build something yourself and you will know.
Related work
While this site was mid-build, I ran a harder experiment on the same codebase: handing Anthropic's Fable model one prompt and measuring how much of a complete, production system — a live wildfire tracker for the fire burning nine miles from our house — it could deliver in a single turn. The write-up, including where the one-turn thesis held and where it needed a human, is its own case study: What can Fable accomplish in one turn?