In short The quality of an AI answer depends heavily on how you framed the task. The prompt is that framing. And AI Fluency is Anthropic's frame of four habits (the four "D"s) that turn random luck into a predictably good result.
What a prompt is and why it decides
The model doesn't read minds — it builds an answer from what you wrote (next-token prediction). A vague request → a vague answer: the model fills the gap with its own plausible guess, not what you meant. So the prompt is the main lever, and it's a trainable skill.
What makes a prompt good:
- Specificity. Not "shorter," but "3 sentences." Not "formal," but "a business tone, no slang."
- Context. Who you are, who it's for, the goal, what already exists. The model doesn't know until you say.
- Role. "You are a science-journal editor" tunes both vocabulary and strictness.
- Output format. Table / list / JSON / N points.
- Examples. Showing 1–2 samples of what you want is the most powerful move (few-shot).
- Iteration. Didn't land? Say what exactly was off. That's normal, not a failure.
On the limits of steering (instructions don't guarantee, safety takes priority) — see steerability.
The AI Fluency frame: four "D"s
1. Delegation
Decide what to hand to the AI at all, and what to do yourself: what's the goal? who's better — me, the model, both? how to split the task into pieces the model can handle? The beginner's mistake is handing over whole what should have been split, then being surprised by the mush.
2. Description
How you frame the task — this is prompting: role, context, specifics, format, examples. The more important and complex the task, the more detailed the description.
3. Discernment
Look at the answer critically, not on faith. Is it factually right? Is the reasoning sound? Right format and tone? Smooth confident text ≠ truth (model knowledge and hallucinations). Over time this becomes an eye — you see where the model probably slipped.
4. Diligence
You are responsible for the result, not the model. Check facts and links, don't pass others' work as your own, know where data goes, don't publish the unchecked. Especially when the AI acts on your behalf (Claude Cowork) — verify the plan and the result.
How to apply it
- Delegation: is this even an AI task? split it up.
- Description: role + context + specifics + format + example.
- Answer in → Discernment: check facts, logic, form.
- Off? Refine (iterate) and judge again.
- Diligence: a final check — the responsibility is yours.
Related
The source is the Claude 101 and Introduction to Claude Cowork courses. Related: steerability, model knowledge and hallucinations.