Terminators, Dogs, and Skills
Could your salespeople pass an industry cocktail party test? Could your marketing copy? Inspired by the dogs in Terminator, I built an AI skill to flag writing that sounds like it came from an industry outsider.
I have a test that I used to give my vertical salespeople: if I dropped you into an industry cocktail party, how long could you last before someone figured out you didn’t work in the industry?
The A+ outcome was that, after 20 minutes of talking with (for example) a group of senior publishing executives, someone would ask, “So, Larry, what publisher do you work for?” Larry would reply, “Oh, I don’t work for a publisher. I work at a software company that serves the media industry.”
That was the standard. Larry understood their world well enough that they assumed he belonged to it. He could follow the conversation, ask good questions, talk at a business level, and contribute something useful without revealing through his language or assumptions that he came from somewhere else. That he was an outsider.
Which brings me to dogs and Terminators. Remember the original Terminator movie? In the dystopic future, humans hid underground while machines controlled the surface. Terminators were built to infiltrate those hideouts, and they were getting harder and harder to distinguish from humans. But dogs could tell. When a Terminator showed up, the dogs started barking.
Think of your salesperson as the Terminator and your customers as the dogs. That’s the cocktail party test. How long can your seller talk to them in an informal environment before the dogs start barking?
Woof-sider, woof-sider!
Why does this matter? Because, as a chat with a CMO the other day reminded me, in vertical sales, “nuance is everything.” You need to speak their language. You need to fully understand their problem. “We need to see and then communicate the non-obvious patterns that connect them to each other. You need to be able to use reference customers in the same industry – always critical to vertical sales – even when there are superficial differences, but deeper commonalities, in their use-cases.”
I’ve always said that in vertical sales you want to complete the customers’ sentences when they are describing their problem. You want the customer to think, “heck, these people not only understand my problem, they might understand it better than I do.” You win by demonstrating understanding of the problem.
Let’s think about how one fails the cocktail party test:
- Some people get busted before they open their mouths. The clothes, the look, something gives them away.
- Others get busted during introductions. At a DoD or IC event someone might introduce themselves simply as "Mike." Knowing not to ask their surname is part of knowing the room.
- Some get busted by their vocabulary. They use an unfamiliar term, misuse a familiar one, or sound like a software person talking about healthcare instead of a healthcare person.
- Some miss the social context. They don’t know where people gather, what they read, or whom they listen to. Or they cite someone as an authority whom the people in the room don’t respect.
- Others get the priorities wrong. They enthusiastically explain how to solve a problem that nobody particularly cares about, while overlooking the problem that keeps everyone up at night.
But then my CMO friend said something that surprised me: “and that’s why I still want to review every piece of industry copy that we produce.” And that’s where skills entered the conversation.
Getting a salesperson through the cocktail party test takes hiring, training, and practice. But reviewing written copy suggested an AI solution: building a skill that flags passages likely to make the dogs bark.
Pretty much everybody these days uses AI skills for core messaging, corporate voice, and style – and I love skills for these use-cases. I spent many hours back in the day writing style guides that I knew would mostly gather dust. So when I saw that skills could bring those moldy reference guides to life, I was thrilled.
But this time, for fun, I wanted to build the opposite: an outsider detector.
A dog that would bark when presented with communications that sounded like an outsider. So I spent a few hours, using a combination of Claude and ChatGPT, to build this skill that detects tech outsiders. When presented with a proposed piece of communications, the skill will generate an assessment like this:

Building the Dog
I started by telling the models what I thought signaled tech fluency. Mathematical language was an obvious example: orthogonal, priors, variance, standard deviations, models, weights, and parameters. Sometimes people are doing mathematics. Sometimes they’re describing how they think.
“I need to reset my weights” is a trendy way of saying that you need to reconsider the importance you’ve assigned to different factors. Lila Shroff’s Atlantic article, Why Is Everyone in Silicon Valley Talking Like That?, explores this borrowing of AI language for everyday thought and provided additional examples.
Then there’s the startup and venture dialect: founders, rounds, runway, dilution, Sequoia, a16z. In marketing, it’s ABM, PLG, ICP, and outcomes. But the room matters. “Operators” sounds natural when investors discuss the people running their portfolio companies. It can sound rather different when addressing those companies’ employees.
My initial instinct was that an outsider might reveal themselves by not using this language. But that’s too simple. An insider can write in plain English, and an outsider can learn the vocabulary. We needed to look at the understanding underneath it. Consider these two invented passages:
- We signed a three-year subscription at $300K per year, adding $900K of ARR.
- The pilot worked, but the purchase stalled. The department had budget, security owned the exception, and nobody had agreed who would run the service after launch.
The first uses the vocabulary but confuses annual recurring revenue with total contract value. The second demonstrates understanding of how enterprise purchases get stuck, without relying on much jargon. Adding “orthogonal” to the first passage won’t save it.
So the skill needed to examine language, assumptions, priorities, and operating knowledge. It also needed to explain its reaction. “Sounds like an outsider” isn’t particularly useful feedback. “You’ve assumed that a successful pilot means the organization is ready to buy” gives the writer something to work on.
Trying It on Real Writing
I then tried it against several pieces of writing. These were early checks, not a scientific evaluation. Some samples contained bylines or references that made the author obvious. Nevertheless, the exercise helped us refine what we were looking for.
- Paul Graham. It recognized strong founder/investor fluency. Growth rates and compounding were how the writer approached the problem. The mathematics was part of the argument.
- Martin Casado. It picked up operating details and recurring traps: premature sales hiring creates pressure on engineering; a customer can pay for access to your expertise without becoming a repeatable product customer; a channel partner won’t necessarily create a market for you.
- Benedict Evans. It noticed the “10 meg spreadsheet running a department” and the small team using Google Sheets because the corporate system was too inflexible. These are observations that suggest familiarity with how companies actually function.
- Ezra Klein. His AI essay demonstrated considerable knowledge of the people and debates. But to my ear, it read more like someone explaining the industry from outside it. The skill initially returned mixed signals, which prompted a discussion about the difference between being knowledgeable and sounding native.
The Evans example was particularly useful because it stopped us (i.e., me, ChatGPT, and Claude) from confusing someone’s role with their fluency. An analyst can be native to an industry. We can’t use “observer” as a synonym for “outsider.” (And he was most definitely an insider before he became an observer.)
I also gave it a foreign-policy essay about Japan. It correctly returned “insufficient evidence.” There was no reason to expect tech language in the piece. The absence of insider signals only matters when the subject gives the writer a reason to demonstrate them.
What We Changed
The biggest lesson was to separate three judgments:
- Native fluency. Does this sound like someone who inhabits the reader’s world?
- Subject knowledge. Does the writer understand the concepts and mechanisms?
- Accuracy. Are the specific claims correct?
An insider can be wrong. An outsider can be informed and right. We don’t want the skill to become a fact-checker or, worse, a detector of whether the writer agrees with us. Criticizing the industry doesn’t make someone an outsider any more than praising it makes them an insider.
We also needed to judge the pattern across a piece. It would be easy to excuse a familiar founder’s imprecision while treating a journalist’s awkward metaphor as proof that they didn’t understand technology. The revised skill explicitly excludes credentials as evidence and tells the model to consider alternative explanations before flagging a passage. Explaining an acronym for a general audience, for example, shouldn’t count against the writer.
Finally, we made it look for positive evidence as well as mistakes. Understanding a non-obvious pattern across customers should count for more than knowing which acronym to use. That brings us back to the CMO’s point about nuance. The strongest signal is often an observation that makes the buyer think, “Yes, that happens to us, too.”
Do Try This At Home
You can download the skill as a markdown file, here. You can use it as the basis for your own skill or use it directly. Tell it who the intended readers are: sounding native to a group of VCs is different from sounding native to a group of developers.
I’d begin with copy you already have strong opinions about. Include something that sounds native, something that nearly passes, and something that has no reason to demonstrate industry fluency. Remove bylines and credentials where practical. Pay particular attention to false alarms. A dog that barks at everyone isn’t much help.
For your own vertical, replace the tech examples with distinctions your buyers care about. Ask your customers and best vertical sellers which passages sound wrong and why. Capture the language, assumptions, and missing context behind their reactions. Include examples of writing that gets it right, particularly where it recognizes a problem or pattern that outsiders tend to miss. Then test whether the skill can recognize similar strengths and weaknesses in new copy.
This is also a useful exercise in buyer fluency. Knowing where buyers gather, what they read, and whom they trust helps you decide both what to say and how to reach them. The copy review tests part of that understanding; building the skill forces you to make it explicit.
I wouldn’t use the output as a verdict on a writer. I’d use it to improve the copy and identify gaps in our understanding of the buyer. Sometimes the fix will be a word. Sometimes it will require a customer conversation. Either way, I’d rather hear the barking before we publish.