ai fluency: the 4 skills, and the one you skip.
anthropic just put out a framework for using AI well. it is four skills, and almost everyone only practices one of them.
ai fluency is the ability to work with AI in a way that is, in anthropic's words, "effective, efficient, ethical and safe." anthropic breaks it into four skills, the 4Ds: delegation, description, discernment, and diligence. that is the whole thing. and here is the part worth stopping on: prompting, the thing everyone means when they say "i'm good at AI," is one slice of one of the four.
this matters because of a number i keep coming back to. about 85% of people have access to AI and about 25% use it on real work (see the ai adoption gap). that 60-point gap is not a model problem. it is a fluency problem. people learned to prompt and never learned the other three skills, so the AI stayed a toy.
the four skills, in plain terms
here is the framework, stripped to one line each and pointed at your actual work.
| the skill | what it actually is |
|---|---|
| delegation | deciding what work is yours, what is the AI's, and what you do together. it starts before you open the chat. |
| description | telling the AI clearly what you want, how to get there, and how to behave. this is where prompting lives. |
| discernment | judging whether the output, the reasoning, and the behavior are actually any good. the quality-control skill. |
| diligence | using it responsibly: which tool, who you tell, and owning whatever you ship with your name on it. |
the three ways you work with AI
before the four skills, one distinction anthropic draws that clears up a lot of confusion. there are three modes:
- automation. the AI does a specific task you told it to do. you define it, it executes.
- augmentation. you and the AI think together, back and forth, both shaping the result. this is where most real business value starts.
- agency. you set the AI up to act on your behalf without you in every step. you configure the knowledge and the behavior instead of typing every instruction.
the whole arc of a working AI setup is moving one task from augmentation to reliable automation, then handing the boring parts to agency. that is what "an agent" actually is. not magic. a described, discerned, delegated task you stopped supervising because it earned it.
delegation: the skill that starts before the chat
delegation is deciding what work goes to you, what goes to AI, and what you do together, and it needs two things: knowing your own goal, and knowing what the tool is actually good at. most people skip it and paste a vague ask into the box. the good move is upstream: pick one high-volume, low-risk job you do the same way nine times a week, and hand the AI that. i wrote the picking-the-first-task version in the ai audit, and the full delegation test in what tasks to give ai.
description: prompting is only part of it
description is communicating with the AI, and it has three layers: what you want (the output), how it should get there (the process), and how it should behave (concise or detailed, challenging or supportive). most "prompt tips" only cover the first layer. the leverage is in the other two. i broke the practical version into six moves in how to write better AI prompts. the short version: be specific, show an example of "good," and give it a role.
discernment: the one everyone skips
discernment is critically judging what the AI produced, how it reasoned, and how it behaved. this is the skill nobody puts in the launch post, and it is the one that decides whether AI is safe to rely on. AI produces confident, plausible, wrong answers. that is not a bug you prompt away. it is a permanent feature you manage with judgment.
this is also the fix for the real reason people quit AI: trust. it gave one confident wrong answer near something that mattered, and they got cautious, which is not irrational, it is professional self-preservation. the rule that kills that problem: keep a human on the send. the AI drafts, you ship. nothing goes out with your name on it that you did not read. more on catching the misses in how to tell when ai is confidently wrong.
diligence: the part with your name on it
diligence is using AI responsibly: choosing the right system, being honest about AI's role with the people who need to know, and taking accountability for whatever you use or share. for a business this is not abstract. it is: do you tell the client, are you okay defending this output in a room, and would you sign it. if the answer is no, it does not ship. that is the whole ethics of it. the disclosure version is in should you disclose ai use.
what this changes for you
most people are practicing one skill and wondering why AI is not delivering. here is the reframe:
| the skill people practice | the skills that actually make AI stick |
|---|---|
| better prompts (description, layer one) | picking the right task to give it (delegation) |
| "the perfect prompt" | telling it how to behave, not just what to make (description, layers two and three) |
| trusting the output because it sounds good | catching the confident wrong answer before it ships (discernment) |
| using it quietly and hoping | owning what you ship and disclosing where it counts (diligence) |
the takeaway
you do not have an AI problem. you have a fluency gap, and it is almost always in the same three places: you did not decide what to delegate, you only described the what, and you never built the discernment to catch the misses. the tools are already on your desk. the four skills are what make them work. pick one task, run all four on it, and you are past the 25%.
want the setup, not another framework?
the free system builder walks you through delegation and description on your real first task, so the AI actually does the job and stays. no webinar. you build it in claude today.
get the free system builderor just follow along. new field notes most weeks on x, instagram, and tiktok.