Anthropic released The AI Fluency Index report, where they analyzed nearly 10,000 sessions in Claude. The goal is to measure "AI literacy" based on real dialogues.

We are interested in the section on generating complex artifacts (this accounts for 12.3% of all chats), as it includes code generation.

In such dialogues (compared to regular ones), users manage context much more strictly:
▫️ Clarify the end goal (+14.7%)
▫️ Rigidly fix the output format (+14.5%)
▫️ Feed examples into the prompt (+13.4%)

But the more detailed the input specification, the worse the review on output. As soon as the AI produces a syntactically correct, polished snippet, critical thinking shuts down. Basic fact-checking metrics drop:
▫️ 5.2% less likely to try to find missing context.
▫️ 3.7% less likely to verify facts.
▫️ 3.1% less likely to ask the model to explain the logic behind the solution.

Anthropic explains this as: a well-designed artifact lulls vigilance. And given that AI most often fails on complex tasks, ignoring review at the moment of generation is a questionable strategy.

Based on these metrics, researchers identified three patterns that distinguish advanced users:

1️⃣ Iterativity. In dialogues where the user does not take the first answer but forces the AI to refine the result, the model's logic is challenged 5.6 times more often. The first answer is always just a draft.

2️⃣ Presumption of guilt for the perfect output. If the generated code looks flawless, that's the main moment to pause. This is where you need to ask questions about architecture and edge cases that the model might have ignored.

3️⃣ Meta-context setting. Only in 30% of chats do people set boundaries for how exactly the AI should communicate with them. Adding instructions to the prompt like "Push back if my assumptions are wrong" or "Walk me through your reasoning before giving me the answer" dramatically changes the quality of the result.