14 July 2026

Hello Lab: How I work on my career with an AI — and how this site came out of it

Welcome to the Lab — the part of this site where I show what I tinker with outside of work: AI experiments, IoT and smart home, infrastructure projects and whatever else is on the workbench.

The first post here was originally just going to explain why this site isn’t WordPress. But that would be half the truth. This website isn’t actually the project — it’s the by-product of a much bigger one: for a while now I’ve been working systematically on my own career, together with an AI. Internally I simply call the project MyFuture; my sparring partner is Claude. And since that process has taught me at least as much as any tech topic, I’ll tell both stories here: the what and the how.

The actual project: taking stock after 19 years

After roughly 19 years of working life, things pile up: reference letters, feedback reviews, project documents, certificates, five different versions of a CV. Scattered across folders, mailboxes and drawers. And a problem emerges that many will recognize: the title on paper long stopped telling what you actually did. Anyone reading only my contractual job title learns nothing about platform building, product ownership or leading teams — all of that lives elsewhere, in documents nobody ever lays side by side.

That was the core intention of MyFuture: do a complete inventory, once — evidence-based instead of from memory. Concretely:

Digitize and analyze everything. I collected all my career documents — over 40 of them — and worked through them systematically with Claude: What do they really say? Which statements, numbers and quotes prove which skill?
Build one consolidated career picture. The individual analyses became a central context document: career path, roles, verifiable achievements, verbatim quotes from reference letters and feedback. A single, verified source of truth about myself.
Derive a positioning. Only on that factual basis came the strategy work: What do I stand for? Which roles fit? How do I tell my story so it works for recruiters and engineers?
Build the tools. And only then came the decision to make this website the centerpiece — a place where the short version and the depth can exist side by side, unlike in a two-page PDF résumé.

If you look closely, you’ll spot the pattern: it’s the same approach as my IT Product Selector — gather evidence first, evaluate in a structured way, decide only at the very end. Apparently my favorite pattern, whether it’s about software products or my own career.

What working with Claude actually looks like

“Working with AI” quickly sounds like either magic or laziness — it’s neither. In practice, a typical MyFuture session looks like this:

I bring the material and the decisions, Claude brings structure and stamina. The AI has read every one of my reference letters and can tell you in seconds which quote proves which competency. But whether a quote gets used publicly, whether a phrasing sounds like me, which projects are lighthouses — that’s my call. Every document and every text on this site went through my approval, most of them through several rounds.

Structure before text. Before a single line of final content is written, there’s an outline to approve. It sounds bureaucratic but saves enormously: a beautiful text in the wrong structure is worthless, and structures are ten times faster to discuss than prose.

Brainstorming as dialogue. The most valuable sessions weren’t the ones where Claude produced something, but the ones where it asked the right questions back: Which numbers are you missing for this claim? Why do you leave that position off your CV — the gap shows up in screening. An AI that spots inconsistencies across 40+ documents is an inconvenient sparring partner — which is exactly what makes it useful.

Persistent context instead of Groundhog Day. So that sessions don’t start from zero, we maintain shared working documents: the consolidated career file, concept papers, open-question lists. Claude also keeps notes of its own — for instance the rule to never work directly on the main branch. More on that in a moment.

Why the site still isn’t WordPress

The tech decision was made the same way, by the way: as a deliberate trade-off in dialogue, not out of habit. The short version:

Astro instead of a CMS: Every post is a Markdown file in a Git repo. No admin panel, no database, no plugin updates — git push, done. Also: an AI works brilliantly with a Git repo full of text files; with a WordPress admin panel, not so much.
One source, several outputs: My résumé lives as structured data in the repo. It powers the project timeline on this site, (soon) a CV PDF built by a CI pipeline, and the context for a chatbot.
GDPR by design: No external CDNs, no cookies, no trackers — hence no cookie banner. It’s possible if you build it that way from the start.

What I do well — and what, honestly, I don’t

This is the part that matters most to me, because it’s transferable. After quite a few sessions with Claude, I can say fairly clearly what works and what I learned the hard way.

What has proven itself:

Real material instead of a hallucination surface. I never let the AI imagine things about me — it works exclusively with my real documents. The quality difference is dramatic: “strong communicator and team player” becomes a verifiable quote from an actual feedback review.
Clear guardrails up front. No third-party names, no employer internals, no secrets in the repo, privacy considered from the start — I state rules like these before the work, not afterwards while cleaning up. The AI follows them remarkably reliably — if you actually say them out loud.
Let it verify instead of clicking myself. Claude tests its own changes: starts the dev server, fills in the contact form, checks the build passes, verifies the live URLs after deployment. I look at outcomes, not at every intermediate step.
Keep the decisions, hand over the legwork. The division of labor is always the same: research, drafts, consistency checks, anything repeatable → AI. Judgment, taste, approval, anything with my name on it → me.

What I had to learn (and partly still am):

Worked directly on main for too long. This site auto-deploys on every push. Which is wonderfully convenient — until a leftover test heading, complete with a private greeting line, gets published along with everything else and sits on the public internet for days. That’s exactly what happened with the first version of this article. The lesson: feature branches and pull requests, even on a one-person project. Claude has since written that rule into its own notes and now creates branches unprompted.
Dictated prompts are a game of telephone. I speak many of my instructions via voice recognition — which occasionally turns “Astro” into “astronomer” and “brainstorming” into something rather creative. Claude almost always guesses what I meant. But almost always means: now and then a round goes in the wrong direction that a more precise request would have saved.
Too many topics in one request. My requests like to bundle four concerns into one paragraph: wire up the form, a new email address, a privacy worry and a hosting question. It works — but cleanly separated requests get done faster and more thoroughly. That discipline has to come from me, not from the AI.
Ask for the “why”, not just the “what”. When the contact form suddenly failed in the cloud environment, the cause was a concept I hadn’t had on my radar: separate environment variables for preview and production deployments. The AI would have explained it to me during setup — had I asked. Since then I add “briefly explain why” more often. It costs two minutes and saves the debugging afterwards.

If I had to boil it down to one sentence: AI doesn’t replace thinking — it punishes non-thinking faster. Vague instructions, missing guardrails, skipped processes — all of it comes back at you at AI speed. Given clear material, clear rules and clear decisions, though, it’s the most tireless colleague I’ve ever had.

What’s next

The plan for the Lab: building the CV pipeline (résumé as code, PDF from CI), an “ask my résumé” chatbot backed by an EU-hosted model — and eventually a small Kubernetes cluster as a playground. All by-products of the same project: MyFuture, built together with an AI, decided by a human.

← Back to the lab