Understand · 10 minutes

The four competencies of AI fluency

Working with AI effectively, and safely enough to put your name to the result. The 4D framework splits that into four competencies: deciding what to hand over, saying what you want, judging the answer, answering for the result. The prompt is the second one.

How they fit

Before, during, after.

A chat history shows only the short loop. That is why the test also asks about the before and the after.

Delegation, Description, Discernment, Diligence: two nested loops LONG LOOP · AROUND THE CONVERSATION BEFORE Delegation AFTER Diligence DURING · SHORT LOOP Description Discernment describe · judge · describe again what you answer for changes what you delegate next time Quality is decided in the short loop. Risk is carried by the long one.
A question for a one-to-one

“On your last real piece of work with AI, what did the second attempt change?”

Three contexts

Three ways of working

The loop runs differently in each. An AI project belongs to one of the three; the team view describes its projects this way.

The framework

Twelve behaviours

Three observable behaviours per competency, the sub-competencies. The definitions in quotes are the framework’s own.

The scale

Four levels

The original framework has no levels. This site adds four, read with four rules.

Observable

A level is read from what a third person could see in the work.

Volume does not count

Heavy use at level 1 stays level 1.

Open to every job

The top level means writing down a practice that others reuse. It does not require managing anyone.

Hard tasks tell levels apart

On an easy task everyone looks at ease. Levels show when the answer disappoints.

The grid

Forty-eight descriptors

Each cell is one descriptor: a sentence you could check on real work. Choose one to see a situation from the test at that level, and what the next level asks.

Descriptor and example

Choose a cell of the grid. The example changes with the setting chosen here.

Why the top level cannot be self-declared
Every level-4 cell ends the same way: and others use it. Levels 1 to 3 describe what a person does; level 4 describes what happened to what they made, so the evidence sits with other people. In a team where nothing circulates, nobody can reach level 4, and that says something about the organisation.

What is already measured

Checking the facts is the rarest behaviour.

Anthropic’s AI Fluency Index: 9,830 real conversations on one tool, one week of January 2026. It measures conversations, not people, on eleven of the framework’s twenty-four indicators, the ones a conversation shows; none of them belongs to Diligence.

Sources: The AI Fluency Index (2026, 9,830 conversations, 20–26 January) · The 4 Ds, behavioural indicators · AI Fluency Framework

Two minutes

Four questions

One per competency. Answer with your last real piece of work in mind.

Take the testSee the team view