I recently got introduced to Athletic Brewing Co Non-Alcoholic Beer at a dinner with friends.
I was skeptical. I had tried non-alcoholic beer before, and most of it tasted like someone had described beer to a glass of sparkling water. But I had a long drive home that night, so I gave it a shot.
Oh boy. They did it right.
It tastes remarkably like a real can of cold beer, minus the hazy after-effect. Apparently, this is something I have been looking for without realizing it.
As a dad of a two-year-old, I often find that the most productive part of my day starts after she goes to bed. She does not go to bed particularly early, which is a separate problem for another day.
By then, I have already spent a full working day draining my mental battery, followed by a few hours serving as the safety officer for a monkey-bar enthusiast. And yes, by “monkey-bar enthusiast,” I mean my daughter.
So I really enjoy those last two hours of the night.
Laptop open. Cold beer next to me. Tinkering with some random idea that probably did not need to exist.
There is only one problem.
Maybe it is my disordered sleep schedule. Maybe I am simply getting older. But even one light beer can dull my brain enough that a long ChatGPT or Claude response starts looking suspiciously like a terms-of-service agreement.
I call this CUI: Coding Under Influence.
In my twenties, I could drink a bottle of wine while debugging something at 2 a.m. Apparently that superpower has expired.
So a good non-alcoholic beer solves one half of the problem.
The other half is: can I make the AI easier to understand when my brain is operating at 70%?
And gosh, while we are at it, please stop with the em dashes. And the endless “it’s not X, it’s Y” constructions.
They can keep “you’re absolutely right,” though.
Who does not need a little self-assurance in life? It is the least AI can do for me.
Enter the airplane maintenance manual
Thanks to modern recommendation algorithms, and more importantly my enthusiastic consent to continuous privacy invasion, YouTube recently dropped a video into my feed.
It appeared next to a mechanical keyboard recommendation, which appeared next to the latest Fireship video.
The algorithm knows me disturbingly well.
The video introduced me to Simplified Technical English, or ASD-STE100.
STE is a controlled form of English originally developed for aerospace maintenance documentation. Airlines had a practical problem: maintenance manuals were written in English, but many of the people reading them were not native English speakers.
When you are repairing an airplane, ambiguity is not charming.
So STE created a set of rules designed to make technical instructions extremely difficult to misunderstand.
And once I started reading those rules, I immediately thought:
Wait. I want my AI to talk like this.
Not all the time. I do not want Claude writing my feature ramp email like a Boeing maintenance manual.
But when I am debugging something, learning an unfamiliar system, or asking an agent what I should do next, this style is kind of perfect.
Here are a few rules that jumped out at me.
1. One word should have one meaning
Technical writing loves synonyms.
“Check the configuration.”
“Verify the configuration.”
“Validate the configuration.”
“Confirm the configuration.”
Do these all mean exactly the same thing?
Maybe. Maybe not.
STE tries to control vocabulary so the same word carries the same meaning consistently.
For example, instead of casually rotating between start, begin, commence, and initiate, choose one approved term and keep using it.
For an AI response, I would much rather read:
Start the service.
than have the model suddenly decide that the next step is to:
Initiate execution of the service process.
Thank you, Shakespeare. Start the service.
2. Keep sentences short
For procedures, STE limits a sentence to 20 words. Descriptive sentences get 25.
This is particularly useful for AI because AI loves the sentence that contains the answer, three qualifications, an exception, a recommendation, and a philosophical reflection on distributed systems. (like this exact sentence written by ChatGPT)
Instead of:
You can restart the service after updating the configuration, although you may want to verify the environment variables first if you changed the deployment settings.
Give me:
If you changed the deployment settings, verify the environment variables.
Then restart the service.
My CUI brain approves.
3. Put the condition before the action
This one might be my favorite.
Do not write:
Set the switch to NORMAL when the light comes on.
Write:
When the light comes on, set the switch to NORMAL.
That ordering matters.
Imagine I am scanning instructions instead of carefully reading every word. If the action comes first, I may start doing it before I reach the condition.
Funny enough, I encountered almost exactly this situation earlier today while replacing a Korky toilet fill valve.
The package came with a tamper-proof accessory.
The instructions effectively said:
For plumbers and property managers only: install the tamper-proof accessory to prevent tenant or guest tampering.
Perfect.
Now imagine the instruction had been written like this:
Install the tamper-proof accessory by doing X, Y, and Z if you are a landlord or property manager.
There is a decent chance I would have installed the thing before reaching the “if.”
Congratulations. Now I have protected my toilet from myself.
4. Prefer simple verb tenses
STE also avoids unnecessarily complicated verb constructions.
Instead of:
The operator has adjusted the linkage.
Prefer:
The operator adjusted the linkage.
For AI output, this sounds almost trivial. But once you notice how often models wrap simple facts inside layers of tense, passive voice, and qualification, you start appreciating the rule.
Compare:
The configuration has been successfully updated and should now have taken effect.
with:
The configuration changed.
Or, if the agent actually knows:
I updated the configuration.
Much better.
Why this works surprisingly well for AI
The funny thing is that STE was designed partly for people like me.
Not specifically sleep-deprived dads drinking fake beer while talking to Claude at midnight.
But people who need to understand technical English even when English is not their native language.
That is me.
And LLMs have created a new version of the same problem. They can produce essentially unlimited prose at essentially zero marginal effort.
That means the scarce resource is no longer writing.
It is my attention while reading it.
So now I have two tools for late-night tinkering:
Athletic Brewing keeps the alcohol out of my brain.
STE keeps the unnecessary words out of my agent.
A surprisingly good combination.
@AthleticBrewing, I am happy to discuss sponsorship opportunities for future posts of this extremely specific genre.
Try it yourself
There are two easy ways to experiment with this.
The simplest is to add something like this to your prompt:
Use an ASD-STE100-flavored style to respond
The second option is to use an agent skill.
There are now several open-source projects that package STE-inspired writing rules for Claude Code, Codex, Cursor, and other coding agents.
Which, frankly, is exactly what I wanted. GitHub
Just use it selectively.
Runbooks? Great.
Debugging instructions? Great.
Architecture explanations? Probably great.
Your next dating profile?
Maybe keep some ambiguity.