Going Beyond ASD-STE100 - by Quinten Lisowe - The Lisowe

The Lisowe

SubscribeSign in

Going Beyond ASD-STE100

Quinten Lisowe<br>Aug 19, 2026

Share

Ever work with AI and get hit with something like this?<br>You’re absolutely right that the race condition in the shared cache layer is the load-bearing failure mode here — honesty compels me to note that everything else is downstream of it. I want to sit with that for a moment. Both can be true at once: the eventual-consistency model is a feature not a bug and it’s creating these subtle timing windows that only surface under load. We’re going to have to eat this elephant one bite at a time. One thing to gently push back on is the instinct to throw more retries at it; that just papers over the real seam. The smoking gun is the missing fence around the write path, and once we add a proper belt-and-suspenders check there (atomic compare-and-swap plus a short lease), the whole system stops thrashing. That’s not nothing — that’s actually everything. Genuinely interesting how turtles-all-the-way-down the dependency chain turns out to be: fix the cache invalidation and suddenly the downstream queue back-pressure, the retry storms, and the cascading timeouts all quiet down. I now have the full picture, and the unlock is treating that single critical section as the true load-bearing wall of the architecture.

Just endless streams of verbose crap. When doing deep work, clarity of thought and communication is paramount.<br>ASD-STE100 and Flesch Reading Ease

Many people have recently taken to adding the phrase “Respond using ASD-STE100 format” to their AGENTS.md. ASD-STE100 is short for Aerospace, Security, Defense - Standard Technical English 100 (the 100 being a tribute to the older ATA 100 standard). This is a controlled natural language (CNL) designed by the European Association of Aerospace Industries in the 1980’s to help make their training manuals more consistent and readable. A copy of the full 434 page spec can be found here.<br>I’ve found this to be marginally helpful, but there are a couple issues with it. First is that it was designed for the aerospace, security, and defense industries, so the vocabulary is going to be extremely limiting. If it’s used outside of those industries the chances of it being compliant are extremely low, at which point why specify a standard if the model isn’t even going to follow it? Second is that it isn’t adjustable to expertise or fatigue.<br>I prefer an alternative that accounts for things like this. The concept comes courtesy of roon, and is responsible for the highest boost in clarity since “think step by step”:

Link<br>The Flesch-Kincaid Reading Level is the calculation that determines how hard a piece of text is to understand and assigns it a grade level. For example, a sentence written at a 6th grade reading level would read as such:<br>Plants need sunlight to make food, but too little light can slowly weaken their stems and fade their leaves.

The same sentence written at a 16th grade reading level (so a college senior) would read like this:<br>Plants convert light into chemical energy through photosynthesis, but long periods without light gradually weaken growth, reduce leaf pigment, and may kill the organism.

What’s interesting about this calculation though is that it’s just fitting a curve to a set of data points, so the range is unbounded. Here is the same sentence at a -2nd grade reading level:<br>Plants need sun. No sun makes them weak.

And at a 25th grade reading level:<br>Although vegetation uses illumination to transform water and carbon dioxide into stored chemical energy, a prolonged deprivation of sunlight can inhibit photosynthesis, reduce chlorophyll, weaken the stem, pale the leaves, halt normal growth, and eventually cause broad tissue damage that the plant can no longer repair.

The concept of a -2nd or 25th grade is a bit absurd, and I don’t like how this system frames reading complexity as being bounded by the grade ranges of the American school system. I prefer a prior version of the Flesch-Kincaid Reading Level called the Flesch Reading Ease score, which has a normal range of 1-100, where the higher the number is, the easier the text is to read. Not to be misleading, it is still a curve fit, but it does not begin with the pretense that reading complexity is limited to arbitrary grade levels. It also offers a more gradual method of tuning the reading complexity threshold.

I tend to keep explanations somewhere between 40 and 80, depending on how engaged or fatigued I am feeling. In the mornings I will usually start off with 40 when I am wide awake and refreshed, and in the late afternoon switch to something a little higher.<br>I wish I could say that this solves all of my slop problems, but unfortunately it doesn’t. I am still continuously editing, by hand, my AGENTS.md folder. I have found this to lead to higher quality outputs and responses. At some point I will have to go through my skills and rewrite them as well.<br>The...