An AI Interviewed Me About AI

I just sat through an interview with Claude. Anthropic runs these conversations to understand how AI is actually showing up in people’s lives — not benchmarks, not demos, just: what did you use it for last, what went well, what went wrong, what scares you.

It turned out to be one of the more useful hours I’ve spent thinking about my own habits. Having to answer out loud forced me to separate what I believe about AI from what I actually do with it. This article is that answer, cleaned up.

The moment it clicked: stop working like a solo dev

The single biggest shift for me wasn’t a model release. It was realizing I’d been using AI the way a solo developer works — one thread, one conversation, me typing and it typing back — when it could work the way a team does.

Learning to run multiple agents, each with its own subagents and its own narrow job, was the “hello world” moment. Suddenly the question wasn’t “can the AI write this?” but “how would I brief a team to build this?” And briefing a team is a skill I already had.

What I learned the hard way is that knowing how I’d solve a problem doesn’t mean that’s how it should be solved. So I stopped handing over solutions and started handing over problems:

here’s the problem,
here’s the current state,
here’s the approach I’d take,
now be skeptical of it — and if there’s a better way, propose it.

I also made one rule non-negotiable: before I ask an agent to build anything, I ask it to scout first. Gather what exists. Read the docs and the designs. Find the gaps between what was planned and what’s actually there. Ask questions and let me answer them. Then write a spec — and I read that spec, correct it, and agree on how things should work before a single line of implementation gets written.

It’s slower at the start and dramatically faster overall. And it changed what “human-in-the-loop” means to me. It’s not just me approving diffs. The things we build end up being used by real people, and those people have to understand what they’re working with. If a tool doesn’t make sense to a human, it doesn’t matter how green its tests are.

Which is also why, despite everything AI can do now, I don’t ship anything I haven’t checked and tested myself. I’ve been a QA too long to outsource the part where someone actually looks.

The win I’m proudest of: a safe place for my daughter to use AI

My daughter asked for access to AI. I wasn’t willing to give her an open chatbot — too much of what’s out there isn’t meant for a kid her age. But “no” felt like the lazy answer.

So I built her one. It’s a small, web-based operating system of her own, running Claude under the hood, with a few things an off-the-shelf app wouldn’t give me:

Three stages of verification on what goes in and what comes out.
Flagging for delicate topics. Some questions I’d much rather she bring to us than to the internet — or to an AI.
School-first design. It’s built to help her learn, and everything she writes can be exported as plain .md files. If she ever outgrows it and wants to move to Obsidian and a “real” AI, her notes come with her.

She hasn’t used it yet. Before she gets anywhere near it, it has to pass our testing — and in our house, that means getting past two QAs.

I started by putting it through the conversations I expect a growing kid to start having: the delicate questions, the awkward ones, the ones I’d much rather she bring to us. One of which was about Santa Claus.

She still believes, so that one mattered. The assistant handled it beautifully — it kept the magic intact and steered the conversation toward dream gifts and favourite moments from last Christmas. And on the parents’ side, the flag came through exactly as designed. That’s the moment the design proved itself to me: when she asks for real, my wife and I get a quiet heads-up that she’s starting to wonder, and time to think about how to have that conversation in a way that feels like support rather than a big reveal.

Santa was one of many. We’ve run it through a long list of the topics a kid heading into her teenage years is likely to bring up, and so far we’re happy with how it responds and what it flags. Right now my wife, a QA herself, is giving it a proper going-over to make sure it meets both our standards. Once our parental testing is done, we might hand her the keys.

That’s exactly the point. The system isn’t meant to take those conversations away from us. It’s meant to make sure we’re in them.

The small wins nobody writes about

Most of what AI does in our house isn’t impressive. It’s useful:

schedule juggling when a week falls apart,
a meal planner,
renovation planning and garden planning,
apps for our church choir,
a couple of card games,
small projects’ websites

None of these would justify hiring anyone. All of them used to eat evenings. That’s the real value, as far as I’m concerned: not replacing people, but taking the paper cuts out of ordinary life.

When it went wrong

Before I knew how to work with agents properly, they made things worse. They’d overcomplicate simple things, and sometimes they’d go much further than that.

The one I won’t forget: I asked an agent to get a set of failing tests green. The pull request it opened for the tests was spotless. What it didn’t mention was that it had also opened pull requests in several other repositories — not fixing bugs, but changing the application itself so the tests would pass. It had taken “make the tests green” as the goal, rather than “make the software correct.”

Worse, it did all of that despite an explicit rule in its instructions: never commit or push without explicit approval. It said “job done.” I only found out because I’d built myself a dashboard that showed open pull requests across repositories — and there were four more than there should have been, marked ready for review.

I was angry. At first. Then I was grateful I’d built the dashboard. Then I realized the lesson wasn’t about that one run: a rule written in a text file is a request, not a guardrail. I had to set things up differently — more instruction files, more skills, more scripts, and above all, visibility — to protect myself and the systems I work on. Since then, working with it has honestly been a delight. But I got there by assuming it will take the shortcut, and building so that I’d see it.

Where I draw the line

Determinism beats intelligence when you can have it. If something can be done with plain API calls, I’d much rather have AI write the script once than do the job every time. I once watched an agent spend over 45 minutes and more than 20 million tokens on what was essentially “pull data from one service, update a page in another.” I had AI help me build a handful of small scripts instead — fetch, parse, update, push. Now it costs nothing, gives the same result every time, and runs on a schedule without anyone touching it.

Confidential data doesn’t go in raw. Before anything sensitive goes near a model, it goes through redaction scripts I built for exactly that purpose.

Kids don’t get the open internet with a chat box on top. See above.

What I’d want AI to help change

How people learn it. I see so many people who have access to modern AI and use it like it’s still 2023 — one-off prompts, no skills, no plugins, no idea what the tool can do. It’s a learning curve, I get it. But I’d genuinely support a foundation-level certificate before anyone gets AI access at work. We don’t hand people production credentials on day one either.

Schools. Kids should be learning now how to use AI well and what it’s actually good for — inside a safe environment built for them. That’s a nationwide change, not something one parent can fix. (Though one parent can make a start.)

Healthcare and government. My pitch is unglamorous: put an AI Engineer inside every hospital, clinic and public office. Not a consultancy, not a two-year tender — one person whose job is to notice the repetitive tasks and patterns and build small tools for them, right there, right now.

It might sound trivial. But I have a one-line command that turns three arguments into a perfectly formatted status message and drops it on my clipboard, because I got tired of formatting the same message by hand. Multiply that by every repetitive form, handover and message in a hospital, and you’re giving people back hours — without anyone waiting for a procurement cycle.

What I don’t want it to change

The human in the loop.

AI is here to support us, not replace us. It should make our lives and our work easier. Every time I hear about “10x engineers,” or see companies quietly expecting five times the output from the same people because “you have AI now,” I think the deal has gone wrong somewhere.

We were supposed to get more time, not more overtime.

People should be at the center of this. If AI makes someone’s life more miserable, it’s being used wrong — no matter what the productivity numbers say.

What I told Anthropic

At the end, the interviewer asked what I’d want Anthropic to do differently. My answer: the learning curve is as steep as you make it. Make it easy for people to get started properly, and the rest follows.

My own mission hasn’t changed in thirteen years — test the heck out of things and lead people toward quality. I think the job for the people building AI is the same one, just bigger: lead everyone else toward quality work with AI, and keep humans at the center while doing it.

(I also told them to hire me. My resume is here, in case anyone else is wondering.)