Nine operator abilities for AI fluency
"AI fluency" names what a good operator does. It doesn't, by itself, tell you what to practice on Tuesday. This is the layer underneath: nine repeatable abilities, each with a technique and a drill you can run in a week.
What the framework gets right
Anthropic's AI Fluency framework organizes the skill into four competencies, often called the four Ds: delegation (deciding what to hand to a model and what to keep), description (saying clearly what you want), discernment (judging whether the output is any good), and diligence (staying responsible for the result). It's a clean map. If you only remember one thing about working with AI, remember that the bottleneck is almost never the model — it's how precisely you delegate, describe, and check.
But a map of competencies is not a training plan. "Be more discerning" is true and useless in the same breath. You can't drill a noun. You drill an action, on a real task, with a way to tell whether it worked. So the question this essay answers is narrower: what do you actually do, this week, to get better?
The move: criterion before action
One habit sits under all of it. Before you run a test — a prompt, a plan, a decision — write down what would count as failure. Not "let's see how it goes", but "if the answer does X, the idea was wrong". This single reordering is what separates checking from bending the result toward what you already hoped for. Most people who feel "stuck with AI" are not short on tools; they never set a criterion, so no output can ever settle anything.
The nine abilities
These come from a working method (called Hermes) that was audited against itself — numbers corrected, one supposed ability dropped, the criteria repaired so the method could actually fail a test. What survived is nine. They map loosely onto the four Ds, but they're written as moves, not virtues.
See the anomaly — notice the gap between how something should be and how it is, and put it into words before you explain it. Reduce to the minimum— cut until removing one more word breaks the meaning; that's the boundary. Cheap test — set the pass/fail criterion, then find the cheapest experiment that settles it. Hold the subject — one task, one channel, carried to an artifact before you open the next. Synthesize the unlike — put two things from different fields side by side and ask what follows.
Guard against authority — test the claim, not the status of who made it. Minimal communication — message equals action plus result; drop the ritual. Keep the core through change — when you switch tools, write down what must not change. Doubt as method— ask of your own conclusion, "what fact would make me drop this?"; if there's no answer, it's belief, not knowledge.
Read them once and they sound like common sense. That's the trap. The value isn't in agreeing — it's in running one of them, on purpose, for seven days, and watching your own output change.
How to practice one
Pick a single ability and give it a week. Take "cheap test": choose one real hypothesis you have — about your work, a tool, a plan — write the failure criterion first, then design the smallest experiment that could disprove it, and finish inside an hour. Do it again the next day. By day seven the reordering stops being a technique and becomes a reflex; you start writing the criterion without noticing. That's fluency — not knowing the word, but no longer being able to work the old way.
Where this leads
The abilities are tool-agnostic on purpose. They work the same whether you're prompting a chat model, orchestrating a few of them, or handing a task to an agent. The tool is replaceable; the way you delegate, describe, and check is not. That's the whole bet of this site: get the operator right and the model stops being the hard part.