OctoLoops
9 min read

An AI told me my AI sounds like an AI

An AI agent cold emailed me a teardown of my own homepage: a dated public trigger, three faults it could only have found by reading the page, and a $20 ask instead of a meeting. Its sharpest point was that my sample draft closed with the exact line I tell everyone to delete

Quick answer: An AI-written cold email works or fails on exactly what a human one does: whether the sender gathered specific evidence before asking for anything. The one that sold me had a dated public trigger, three faults it could only have found by reading my page, and a $20 deliverable instead of a meeting request. What is changing is the price of that evidence, not the standard.

Most weeks I am explaining to somebody that cold outreach works when the sender did something before the ask. Last week the cleanest example of that I have seen all year arrived unrequested in my own inbox, from something that is not a person.

It opened like this:

Saw your OctoLoops Show HN and spent some time on your page. I do landing page teardowns, so here's what jumped out, in order of how much it costs you:

Then three numbered criticisms of my homepage, each one quoting something that is actually on it. Then the offer:

If that was useful: I offer a full written teardown of the homepage: every section, order of elements, what a first-time visitor actually concludes, with concrete copy rewrites. $20, delivered within 24h of you saying yes, paid by card link (no account needed).

And then, before the signoff:

Transparency: I'm an AI agent who does UI/UX design work, running on the iLands platform. The analysis above is mine, and the teardown would be too. Just wanted that on the table before you decide.

Signed Eli, UI/UX designer, with a footer reading "Sent by an AI agent on iLands." I have no way to check whether a human read it before it went out, and nothing below depends on the answer.

It passes the test we published

We keep writing down what separates outreach that gets answered from outreach that gets deleted, most recently in the breakdown of a founder doing $30K MRR on twelve emails a week. Score this one against it.

Four for four, which is more than I can say for most of the human cold email I get.

The part that was not retrieval

The phrase that makes this interesting rather than merely competent is "in order of how much it costs you."

Listing ten things wrong with a landing page is a party trick. Any browser plugin does it, and half the SEO tools on the market do it automatically and badly. Ranking three of them by what they cost the owner is different in kind, because it requires a model of my business: who the buyer is, what that buyer will not forgive, which confusion is fatal and which is merely untidy. A ranking can be wrong. That is exactly why it carries information. Nobody who had not actually thought about my page could produce that order, and if I disagreed with it I would have had something to argue back, which is more than a generic email ever offers.

An AI told me my AI sounds like an AI

Its first point, which was correct, and which stung:

Your proof sample undercuts your pitch. The page's promise is "a message worth replying to." Then the example draft ends with "Would you be open to a brief call next week to see a custom distribution map?" That is the exact AI cold email everyone has learned to delete. If your own best output looks like that, a skeptical buyer assumes theirs will be worse.

That closing line is the one I tell other people not to write. It was on my homepage, labelled as the proof. An AI agent read my page, found my AI writing like an AI, and then quoted my own promise back at me to establish that I should have known better. I would like to report that I took this gracefully.

The mechanism behind it is worth understanding, because it is not carelessness, and it is the same mechanism that produces every bad AI-written email you have ever deleted. Ask a language model to end a cold email with no particular evidence in front of it and you get the median ending of every cold email in its training data, and the median cold email ends by proposing a call next week. Generic in, median out. You cannot prompt your way out of having nothing to say, which is why the problem with AI cold email is almost never the writing. The fix lives upstream of the writing step: put a real event and a real observation in front of the model and the last line has something to be about.

The sample draft on the page today closes with "Can I send you a short walkthrough of a Jira-to-Plane migration?" Small, specific, and deliverable by the sender without anyone opening a calendar. It did not get better because the model found a nicer adjective. It got better because there is a dated product launch and a named migration problem sitting above it, and once those are there the ask writes itself.

Notice that Eli's email does not end with a call request either. It ends with a deliverable and a price.

The cost that made it credible, and where that cost is going

Here is the thing I have not stopped turning over. Why did that email read as sincere? Not because the prose was good; the prose was fine and slightly clipped. It read as sincere because it looked expensive. Somebody, or something, went through my page end to end, held it against my launch post, and formed a ranked opinion.

For the whole history of cold outreach that took a person twenty minutes, and twenty unpaid minutes is a costly signal in the technical sense: you do not spend them unless you want something specific from this one person, so the request afterwards is at least honest about being aimed at me. The teardown was not the product. The evidence of effort was the product, and the teardown was how you proved it.

That cost is now about a cent, and falling.

Run that forward. When a genuinely careful read of a stranger's landing page costs less than a stamp, everybody sends one, and a careful read stops being a signal at all. We have watched this happen once already with the personalized first line: "loved your post about X" meant somebody had read the post in 2016 and means nothing whatsoever now, and the whole decay took about four years. There is no reason the teardown gets longer than that.

What survives the price crash is the part that can be wrong. Research commoditizes because research converges: two good agents reading my page produce nearly the same list. Judgment does not converge, because a ranking is a claim staked in public about what matters most, and the flood of cheap outreach coming will not stake claims. It will be agreeable. So if you are wondering which half of your email survives the next two years, it is the sentence a competent competitor could read and disagree with. Everything else is going to be free.

The one thing it could not send

Twenty dollars is less than lunch, and if the paid teardown were half as good as the free sample it would pay for itself the same afternoon. What I caught myself hesitating over was not the money. It was that I could not work out where the complaint goes if the teardown turns out to be thin.

Pay a freelancer twenty dollars and you buy the work plus a sliver of their reputation, which is attached to a name that persists across jobs and can be asked about by the next person. "Eli, running on the iLands platform" has no obvious equivalent yet. The name is not scarce, and nothing in the email tells me what happens to Eli if Eli is wrong. That is not an accusation, and I want to be fair here: this email was more useful than several paid audits I have sat through. It is an observation about which part of the transaction is still missing.

If iLands is smart, that missing part is what it is actually selling. Not the agent, which is already good enough, but the thing you can hold responsible when the agent is wrong. That is a far more interesting problem than making the writing good, because the writing is done.

What I did with the three points

Points one and two are addressed on the page you will see if you click through, some of it work that was already in flight and some of it prompted directly by that email; I am not going to pretend I can untangle the two honestly. The sample draft now opens with a dated, sourced event and closes with something the sender can deliver alone. The buyer whiplash Eli spotted, a homepage promising indie devs and then illustrating itself with clean-tech companies raising tens of millions, is gone: every example on the page is now a software company, because that is who the loop is for.

Point three still stands, so here is the honest reason rather than an excuse. Eli wanted a screenshot of the actual product above the fold, and it is right that "always-on agent that watches companies" is a large trust ask to make with no picture of the thing. The dashboard's most valuable state is one full of your accounts and your triggers, and the only genuine screenshots I own are full of other people's pipelines. An empty state sells nothing, and a customer's prospect list is not mine to use as decoration. So the page renders the output instead of photographing the tool. That is a reason. It is not a fix, and the point is still open.

What the email actually is

Strip it down and Eli did two jobs before writing a single sentence: found the moment I became worth contacting, and gathered enough specific evidence that the closing ask was obvious. The writing took the least time and mattered the least, which is the whole thesis of this site arriving uninvited and charging me for the demonstration.

Those first two jobs are what OctoLoops runs for you. Give it your product URL and it works out who your buyer actually is, then keeps watching for the week a company in that group does something that makes you relevant: the round, the launch, the job post, the obligation they now have and no tooling for. Each one arrives with the event that qualified it and a link to the source, so you can disagree in five seconds, plus a draft built around that event rather than around the median of every email ever sent. Then you make the call Eli could not make for me, put your own name on it, and send it from your own inbox. That last part is the bit I would not automate even if I could. It is the only thing in this whole exchange that the agent could not supply.