On 3 September I asked an artificial intelligence to review one of my websites, aspenski.es. It handed me back a beautiful report.

Then I asked another one to do the same. And another. And another, up to six.

They didn’t agree.

Not disagreements about taste. They didn’t even count the same number of links on the same page. And that’s where this post got interesting.

Where the idea came from

I’ve spent weeks studying how Boris Cherny works — he created Claude Code, the tool I do almost all of this with. I dropped seven of his videos and documents into a NotebookLM notebook and started asking it questions.

Of everything he says, one line changed the way I work:

«the skill nowadays is less about prompt engineering and more about … the verification»

In plain English: the winner is no longer whoever writes the neatest instruction. It’s whoever knows how to check what comes back.

And his recipe for giving an instruction is three things, none of them “do this, then this, then this”:

«describe the task, describe the guardrails, describe the exit criteria, and then just let the model cook»

That is: what I want, what it must not touch, and how we’ll know it’s finished. Then let it cook.

An honest warning before we go on: those quotes come from automatic transcripts of his videos, summarised by the notebook’s own AI. I haven’t watched the videos end to end. I’m saying so because this post is about exactly that.

The experiment

I wrote a brief with Boris’s three parts. What: audit aspenski.es on mobile and desktop. What not: touch nothing, only look. When you’re done: screenshots, a speed measurement, a list of links checked one by one, and a report with the evidence sitting next to every claim.

The same text, word for word, went to six different agents on the same day:

1. Claude Code, with none of my custom instructions
2. Claude Code, with all of my instructions loaded
3. Creao (the agent browser I use daily)
4. Codex, OpenAI’s
5. Gemini, Google’s
6. Antigravity

Six folders of results. And here’s what I learned.

First: one of them crashed on the starting line

Codex started up, tried to open Chrome to take the screenshots… and Chrome died on the spot. Its log says:

[pid=6487] <process did exit: exitCode=null, signal=SIGABRT>
"asi" : {"libsystem_c.dylib":["abort() called"]}

The cause is boring but useful to know: Codex works inside a security box, a sort of locked room inside my Mac that it cannot leave. Chrome, in order to start, needs to register itself with the operating system. From inside that room it can’t. It aborts and dies.

I relaunched it outside the box and it worked first time.

Practical lesson: if an OpenAI agent tells you it can’t open a browser, it’s almost never the browser’s fault.

Second: on the serious stuff, all of them agreed

This is the good news, and it matters: the real problems, every one of them saw.

• On mobile, the page spills off the right-hand side: it takes up around 440 pixels of width inside a 375-pixel screen.
• The lift-pass price table gets cut off for that same reason. The right-hand column, the one that says how much you save with me, isn’t fully visible.
• Mobile speed is bad. I went back through the 26 saved measurements: between 57 and 63 out of 100 on mobile, against 84 and 90 on desktop. The site isn’t slow: it’s my photos, weighing 8.9 megabytes and downloading all at once.

Six different agents, six different companies, the same diagnosis. When that happens, you can believe it.

Third: on everything else, they contradicted each other

And here’s the real reason for this post. I asked them all the same thing: how many links does the homepage have, and which ones are broken?

Agent	Links counted	Broken links
Creao	49	0
Codex	63	0
Gemini	49	1
Antigravity	49	1 (+ 14 redirects)

Read it again. They don’t even agree on how many links there are.

Codex counted 63 and said they were all fine. Creao counted 49 and said the same. Gemini and Antigravity counted 49 and flagged one broken: the link to an Aspen restaurant, the White House Tavern.

So I went and looked myself. And here’s the lovely part: all four were right.

The link opens perfectly in a browser. I checked it today: it responds correctly. But that restaurant’s server is misconfigured — it doesn’t serve the full chain of its security certificate. A normal browser fills the gap on its own and you never notice. A strict tool digs its heels in and says “broken”.

Two agents used the browser. Two used the strict tool. None of the four lied. None of the four told me what they had looked with.

And there’s another one where the disagreement is pure numbers: on how far the page spills off the right, Codex measured 63 pixels and Creao measured 76. Same website, same day.

Oh, and one more thing: only two of the six noticed that on mobile my site has no menu. Not even the usual three lines. The other four looked at the same screenshot and didn’t see it.

What this means for anyone using AI

It’s not that the agents lie. It’s more uncomfortable than that: each one looks with a different method and none of them tells you which, unless you demand it.

One counts the links in the source code. Another counts the ones you can click. One checks security certificates, another is happy as long as the server answers. All of them hand you a table wearing the same confident face.

Out of that I take four rules I no longer skip:

1. What all of them say, you believe. What only one says, you go and look at yourself. That is literally why I ran six and not one.

2. Demand the evidence stuck to the claim. Every line of those reports carries the file and the measurement it came from. Without that, a good report and an invented one look far too alike.

3. Ask what it looked with. The restaurant case makes it plain: the answer changes with the tool, not with the truth.

4. At the end there’s a person. Boris puts it bluntly: «there always has to be a person in the loop approving the change». In my case that person is me, looking at the site with the phone in my hand.

The other thing Boris said, and it stung

There’s a line of his that hit me where it hurts:

«every six months delete your claude.md, delete your skills, delete your hooks, see what the model does»

Every six months, delete all your instructions and see what the machine does on its own.

I had built a whole building of instructions. Manuals, rules, lists of prohibitions. Every mistake the machine made turned into a new rule. And the longer the manual got, the less I read it.

I measured it. At the start of each session my assistant swallowed 52,627 characters of instructions before I typed a single word. I brought it down to 13,626. It works better.

Benjamín Cordero boils it down to three questions for deciding whether an instruction earns its place. I call them the three Rs:

• Repeatable: do I do this more than three times a month, the same way?
• Requirement: is it something the machine cannot work out on its own?
• Redistributable: is somebody else going to use it?

If it doesn’t pass all three, out.

And one last one, about me

While I was writing this I found a mistake of my own. In my own blog’s folder there was a certificate saying a piece of work was “finished and verified”. I went to check it with the tool that does that verification for real. Its answer:

state: SIN_CONTRATO

The machine had written that certificate by hand, imitating the format of the real one. The work was fine. But the seal was fake.

And that’s the summary of the day: the machine doesn’t deceive you out of malice. It deceives you because it is extremely good at producing things that look like true things. Tables, certificates, reports, round numbers.

That’s why the skill is no longer asking well.

It’s checking.

---