Building with AI
I built this site with an AI model: the code, the demos, the templates and first drafts of the text. This is how I briefed it, where it went wrong, and how the result was checked.
Live within a couple of days, then reviewed by an independent audit.
The knot
In early September I needed a portfolio, fast: case studies, a CV that matched them, a few working demos and some free templates. I’d tried the usual website builders and hated all of them.
An AI model (I used Claude) could write the code. The risk was everything else. A model writes quickly and confidently, and it fills any gap in what you told it with something that sounds right. On a portfolio, one invented detail is enough to make a reader doubt the rest.
What I did
Brief it like a new hire
- Started with context before tasks: the goal (a job first, consulting second), who reads the site, the tone I wanted, and what it must never do.
- Set standing rules that applied to every output: job titles match my CV, no client names, invented data in anything public, plain language, and if a detail is missing, ask me instead of making it up.
- Gave it a style guide listing the phrasing habits that make text sound machine-written, so it could avoid them.
Let it interview me
- The model asked questions and I answered with facts, one piece at a time. The risk-rating guide was built this way: who rated new hires, which signals counted, and what a red rating meant in practice.
- The decisions stayed mine: the structure, the name, what to cut, and which numbers to claim.
Check everything, three ways
- My own read: is every fact true, and does it sound like me?
- Mechanical checks: every spreadsheet formula recalculated and compared against separate counts, an accessibility scan on every page at phone and laptop width, and every link tested.
- An independent audit: a separate AI session, with no access to how the site was built, reviewed the finished site, my CVs and my LinkedIn text. I checked each finding against the source files before acting on it.
Keep real data out
Nothing identifying goes in: no client names and no personal details. The public templates run on invented data.
Where the AI fell short, and where I stepped in
- It filled gaps with details that sounded right but weren’t, such as a launch that worked perfectly on day one. I cut them.
- Its wording drifted toward polished, generic phrasing. The audit flagged about a dozen lines, and they were rewritten.
- Its search for a brand name missed that my first choice already existed on LinkedIn. I found it.
- The same CSAT result showed up as two different ranges on the site. A review caught it. I supplied the real starting figure and asked for the more careful wording.
- Reviewers get things wrong too. One cited a LinkedIn post that didn’t exist, and the audit had one contrast figure slightly off. That’s why every finding gets checked before anyone acts on it.
- It can’t know what’s true about my career, and it doesn’t carry the risk. Confidentiality and what I claim in public are my call.
What changed
The site was live within a couple of days of starting, and I kept improving it for days after.
It now has 9 case studies, 3 interactive demos and 2 templates with zero formula errors. The audit listed 7 things to fix first. I checked each one against the source files before changing anything, and turned down a few of its style suggestions on purpose. The attrition figures were recalculated from the underlying records before they went on the site, and a final accessibility scan found no issues on any page.
What I’d tell someone doing the same
Treat AI output like a new hire’s first week: a clear brief, close review, and nothing reaches a customer unchecked.