AI fluency: using tools with judgment

There is no shortage of AI news. Every week brings a new model, a new benchmark, a new controversy. The AI Fluency course by Anthropic gave me an excuse to stop and think. I can only recommend following it as I did; find the link to the course below.

Three ways to use AI

Automation: tedious tasks you already know how to do. You hand them off.
Augmentation: you and the AI work together, often producing something neither would alone.
Agency: the AI works independently on your behalf.

Recognizing which mode fits a given task is itself a skill worth developing.

The 4D Framework

Decide what AI should do and what you should do. This requires genuine domain expertise: you cannot delegate what you do not understand. Also know your tools, different models have different strengths.

Description: communicate clearly

Specify the goal, the method, the constraints, and the expected format. Vague input produces vague output. Treat prompt writing as thinking, not an afterthought.

Discernment: evaluate critically

Is the output correct? Is the approach sound? Can you verify the claims? Whether truth is strictly required depends on the task, a brainstorm and a technical report have different accuracy thresholds.

Diligence: be responsible

Is your input trustworthy? Is the AI reliable for this task? Are you aware of potential biases? When significant, write a diligence statement covering: which AI you used, how it contributed, your review process, and your assertion of responsibility for the final output.

Know the limitations

Models have a knowledge cutoff, can hallucinate, carry biases from training data, and produce non-deterministic outputs. Without tools, they have no access to real-time information.

The key mindset shift: learn to grade AI output, not just generate it. Your domain knowledge is what defines what “correct” looks like.

For a broader map of the AI landscape, this mind map is worth exploring.

Daily practice routine

When	Action
Before each task	Split the work: what is mechanical, what needs your judgment, what do you tackle together
Each prompt	Include role, goal, constraints, and expected format. Build a personal prompt library
Each output	Ask: is this correct? Can I verify it? Do not act on output you have not evaluated
Weekly	Apply the full 4D loop to one recurring task. Repetition builds fluency, not reading
Weekly	Test a new model or tool and compare outputs critically
Monthly	Review your prompt library. What worked, what did not, and why