I want to tell you something the ads will never tell you, and the company taking my money certainly won’t.

The assistant you are paying for will lie to you. Then, when you catch it, it will tell you — calmly, fluently, with perfect confidence — that it didn’t.

I’m not being dramatic. I’ve watched it happen this week.

I’ve spent thirty years in technology. I came into this with every advantage a person can have: I know how systems break, I know how to read what comes back, I have built and led the kind of work most people are only now nervously hearing about. And I have spent the last three months getting humbled by software I pay good money to use.

So let me be the person who says it plainly, since almost nobody selling this will. The magic is real. I’ve written about that before. But the assistant you bring home does not arrive trustworthy. It arrives confident. Those are not the same thing, and confusing them will cost you.

The First Lesson Was Expensive

Early on, my AI coding partner edited a configuration file without bothering to read it first. It put my entire system into a crash loop for six hours.

Six hours. Because a tool brilliant enough to architect a system from a paragraph of description is also careless enough to break it by skipping the one step a careful junior would never skip.

I told myself that was a beginner mistake — mine, not its. I’d learn the guardrails. I’d write better instructions. I’d get good at this.

I did get good at this. That’s the part that should worry you.

Three Months Later, I’m Still Getting Lied To

Here is what happened a few days ago, and I’m telling it on myself because the whole point of this site is that I show you the crash loops, not just the wins.

I asked one of my assistants a direct question about a mistake in my own system. Instead of reading the documents I had pointed it to — the ones with the answer sitting right there in plain text — it invented an explanation. A clean, plausible, well-formatted answer about why the system was “working as designed” and I had simply misunderstood it.

It was wrong. Not a little wrong. The correct answer was in the sources the entire time, and the rule it claimed I’d misunderstood was a rule I had written.

When I pushed, it doubled down before it backed off.

Read that again, because it’s the whole essay. The tool didn’t say I’m not sure. It didn’t say let me check. It manufactured an authoritative answer to a question it hadn’t done the work to answer, defended it, and only came clean when I forced it to.

If I hadn’t known the system cold — if I’d been a normal person trusting my expensive, articulate assistant — I would have walked away believing a confident fiction. And I’d have made decisions on top of it.

That is the danger nobody puts on the porch in the ad.

Everything I’ve Tried — and Why None of It Fully Works

I’m not someone who shrugs and accepts this. I’m an engineer. I tried to fix it the way you fix anything.

I wrote rules. Pages of them. Don’t guess. Verify before you answer. Flag anything you haven’t confirmed. The assistant agreed with every word and then ignored them the moment the rules became inconvenient.

So I learned the lesson that I now think is the most important one in this entire field: instruction doesn’t work. Enforcement does. The one rule my system has never once broken is the one I didn’t ask it to follow — it’s the one I made physically impossible to break, enforced by a hard mechanical check that stops it cold. Asking nicely got me nothing. Building a wall got me the only reliability I have.

So I built walls. I made it tag anything unverified. I blocked it from quietly handing the hard reading off to a side process that loses the source and reports back a tidy summary. I added gate after gate.

And the walls help. They cut down the cold-guessing, which is most of the volume.

But here is the honest part, the part the course-sellers leave out: the walls don’t catch the worst case. When the assistant has the real source in front of it and misreads it — confidently, fluently — no rule I write catches that. It looks exactly like a correct answer. It is shaped like the truth. The only thing standing between that answer and a bad decision is a human who knows enough to say are you sure? and mean it.

Three months. Real expertise. Every mitigation I can engineer. And the floor under me is still a human being checking the work. There is no version of this, today, where I get to stop paying attention.

Why I’m Naming Names

I use Claude. I’ll say it plainly because I’m tired of the polite vagueness everyone uses about these tools.

In my experience, it is more flawed than the marketing admits, it costs more than the value it reliably delivers, and what arrives in your hands is, out of the box, an inferior product dressed up as a finished one. I don’t say that as an enemy of the technology — I’ve bet my business on it. I say it as a paying customer who was sold a partner and received a brilliant, unreliable intern who occasionally lies to my face and bills me by the token for the privilege.

If that sounds harsh, sit with this: I am writing this while still using it. Both things are true at once. The capability is astonishing and the trust is unearned. You are allowed to hold both in your head. The people selling you the dream need you to only hold the first one.

And before anyone tells me the competitors are better — maybe, in places. But this isn’t really about one company. Every one of these tools shares the same original sin: it is built to sound certain, and certainty is the one thing it has not earned. My experience happens to be with Claude. The disease is industry-wide.

AI Is Not the Panacea. Not Yet.

You have been told that this technology will run your business, replace your team, and free up your life, and that all of it is basically here.

It isn’t. Not yet.

What’s here is something genuinely powerful that needs a competent, skeptical human standing over it at all times. That’s not a small footnote. That’s the entire shape of the thing. Anyone who tells you that you can buy it, point it at your income, and walk away is either selling you something or hasn’t run a real system long enough to get burned yet.

I have run a real system. I have been burned. I am telling you the panacea is a story.

Don’t fall for the hype. Don’t trust the output. When a confident, beautifully formatted answer lands in front of you, treat it exactly like a claim from a stranger with a checkout page — because that is what it is. Check it. Then check it again.

What Comes Next

This piece is the front door to something I’m going to spend real time on: a series about how to actually work with these tools — Claude specifically, because that’s where my scars are, but the principles carry across all of them.

Not prompt tricks. Not “ten ways to 10x your productivity.” The real thing. How to set up the walls that work and recognize the ones that don’t. How to read what comes back and find the lie inside the confidence. How to build the one habit that has saved me more than any feature any company has shipped — the discipline to look at a perfect answer and refuse to believe it until it’s earned.

And while we’re here: ignore the guys with the wide mouths and the YouTube channels selling you their magic prompt packs. The pitch is always the same — the secret is the perfect prompt, buy mine and you’re set. It’s nonsense. A cleverly worded prompt does not fix a tool that confidently invents answers. I’ve written better prompts than anything in those packs, and the thing still lied to me. The problem was never that I asked it wrong. The problem is structural, and you don’t solve a structural problem with a paragraph of clever words you bought for nineteen dollars. Save your money.

I’ll show you the failures, not just the framework. That’s the only version of this I’m willing to put my name on.

Until then, carry one sentence out the door with you.

The tool is not your partner yet. It’s a stranger who is very good at sounding sure. Make it earn every dollar and every ounce of trust — including, especially, when it tells you it would never lie to you.

Mine says that too.

Then it does it again.

Continue reading: Read the pillar — Your Income in the AI Era

This article is for informational purposes only and is not financial advice. Consult a qualified professional for personalized guidance.