1Purpose first: why a framework at all?
Before the mechanism, the reason it exists.
Most people meet AI as a text box and learn tricks: magic phrases, templates, "act as an expert". Tricks expire the moment the model changes. The AI Fluency framework targets the part that does not expire: how a person decides what to hand over, says what they want, checks what came back, and owns the result.
Three failure patterns it is designed to prevent:
- Over-trust. Accepting fluent, confident output that is wrong.
- Mis-delegation. Giving AI work it is bad at, or keeping work it would do better.
- Unowned output. AI-produced work goes out and nobody is accountable for it.
2What AI fluency is
Definition, origin, and what it is not.
AI fluency is the ability to work with AI effectively, efficiently, ethically and safely. It was developed by Prof. Rick Dakan (Ringling College of Art and Design) and Prof. Joseph Feller (University College Cork). Anthropic teaches it in the Academy course AI Fluency: Framework & Foundations, which is what the post is about.
3Three ways to work with AI
The framework describes three modes of interaction. The 4Ds apply in all of them.
The more autonomy the AI has, the more weight lands on two competencies: Delegation (what you allow, decided before it runs) and Diligence (you will not watch every step, so the checks must be built in).
4The 4Ds at a glance
Four competencies, one loop in the middle.
| Competency | The question it answers | Its three parts |
|---|---|---|
| Delegation | Should AI do this, and in which mode? | Goal and task awareness · Platform awareness · Task delegation |
| Description | How do I make my intent unmistakable? | Product · Process · Performance |
| Discernment | Is this actually good? | Product · Process · Performance |
| Diligence | Who answers for it? | Creation · Transparency · Deployment |
5Delegation 🤝
Deciding whether, when and how to involve AI.
Goal and task awareness
know the workBe clear on the objective, then break the work into pieces. Some pieces suit AI, some need human judgment, relationships or accountability.
Platform awareness
know the toolKnow what a given AI system can and cannot do. A model that drafts well may still be unreliable at exact arithmetic or at facts it was never given.
Task delegation
pick the modeFor each piece, choose: automate it, collaborate on it, or let an agent run it.
6Description ✍️
Communicating intent so the AI can act on it.
Product description
what I wantSpecify the output: goal, audience, format, constraints, what to avoid.
Process description
how we get thereGuide a multi-step piece of work through back-and-forth: outline first, then sections, then polish.
Performance description
how it should behaveDefine how an AI should act when it works on its own for other people. This is what a system prompt for an agent is.
Spec versus search query, as in the post:
| Search-query style | Spec style |
|---|---|
summarise this contract risks | Goal, audience, constraints, format and a fallback, all stated |
Goal: List the risks in the attached contract. Audience: A sales lead with no legal background. Constraints: Only use the attached text. Do not give legal advice. Format: Max 5 bullets, each with the clause number. If unsure: Say "not stated in the contract" instead of guessing.
7Discernment 🔍
Judging what comes back. Fluent is not the same as correct.
Product discernment
is the output good?Check accuracy, completeness, bias and fit with your goal. Persuasive wording is not evidence.
Process discernment
is the collaboration working?Notice drift, flattery, or an AI that agrees too easily. Adjust how you work with it.
Performance discernment
is the system serving users?For AI that runs on its own, judge its behavior through tests and real user feedback.
Description and Discernment form a loop. Each round exposes gaps, and the next description closes them.
8Diligence ⚖️
Taking responsibility for what AI helped produce.
Creation diligence
while making itChoose tools and methods thoughtfully, watch for bias, and think about who is affected.
Transparency diligence
when sharing itBe honest with your audience about how much AI shaped the work.
Deployment diligence
before releasing itFact-check and test before the output reaches other people. Once it carries your name, you own it.
9A worked example
Task: prepare a one-page summary of a customer's published annual report before a meeting.
| D | What you do |
|---|---|
| 🤝 Delegation | AI extracts and drafts (augmentation). You decide which points matter for the meeting. Anything about the customer's intent stays your judgment. |
| ✍️ Description | State the audience, the length, the sections you want, and "only use the attached report; say when something is not stated". |
| 🔍 Discernment | Spot-check every number against the source. Look for claims the report never makes. Re-describe where the draft is vague. |
| ⚖️ Diligence | Fix errors before it circulates. If you share it, say it was AI-assisted and checked by you. |
10What is universal and what is specific
What will still hold when the next model arrives.
| Idea | Class | Why |
|---|---|---|
| The 4 competencies and the Description-Discernment loop | Universal | Describe human skills, not any tool |
| Automation, augmentation, agency | Universal | A spectrum of autonomy, true for any AI system |
| The names of the 12 sub-parts | Framework vocabulary | Dakan and Feller's labels; useful shared language, not laws of nature |
| The course and certificate | Platform | Anthropic Academy offering |
| Specific prompting techniques (tags, role lines) | Fastest to age | Model-dependent; they live inside Description and need re-testing per model |
11Try it on your own work
The same four competencies, applied to a customer support assistant as an example.
| D | What it looks like in a support assistant |
|---|---|
| 🤝 Delegation | Let the assistant handle routine intents (billing, technical, general). Anything it cannot answer escalates to a human. That boundary is a delegation decision. |
| ✍️ Description | A classifier prompt that says "reply with ONLY the category name" is a product description. Behaviour rules for the whole assistant are a performance description. |
| 🔍 Discernment | A fixed test set of real questions, scored every time the prompt or model changes. Re-checking your own expected answers is discernment applied to your own labels. |
| ⚖️ Diligence | Validate the model's category in code with a safe default, add retries, and do not publish claims you have not measured. |
12Check yourself
Why is "write better prompts" not the same as AI fluency?
Prompting is skill inside Description only. Fluency also covers what to delegate, how to judge output, and who is accountable.
Which two Ds matter most as AI gets more autonomous?
Delegation and Diligence. You set the limits in advance and build checks in, because you will not watch every step.
What does the Description-Discernment loop add to a one-shot prompt?
It turns a prompt into a process: evaluate, find the gap, re-describe, repeat.
Who is responsible when AI-drafted work goes out under your name?
You are. That is Diligence.