Your AI Fluency Collapses Exactly Where the Code Ships

Anthropic took 9,830 Claude.ai conversations from a single week in January 2026 and classified them against 11 directly observable behaviours drawn from the 4D AI Fluency Framework — Delegation, Description, Discernment, Diligence — built by professors Rick Dakan and Joseph Feller. One result should stop every engineer mid-scroll.

In the 12.3% of conversations that ended in an artifact — code, a document, a working tool — people wrote a much better brief and ran a much worse audit. Stating the goal up front rose 14.7 percentage points. Specifying the output format rose 14.5pp. Handing over an example rose 13.4pp. Then, in those same conversations: flagging missing context fell 5.2pp, checking facts fell 3.7pp, questioning the model's reasoning fell 3.1pp.

Read that again with your own week in mind. The conversations most likely to end up in a pull request are the ones where scrutiny is lowest.

The behaviour that drags everything else with it

Iteration showed up in 85.7% of conversations, and it is not a small lever. Iterative conversations averaged 2.67 fluency behaviours against 1.33 in the ones that stopped at the first answer — literally double. Inside them, people were 5.6x more likely to question the reasoning and 4x more likely to spot missing context.

My rule of thumb: never accept a first draft. Not because first drafts are bad, but because the act of pushing back is what switches the rest of your judgement on.

A second gap is worth your attention: only 30% of conversations contained any instruction about how the person wanted to work with the model. For a dev that is nearly free to fix — a CLAUDE.md, a project instruction block, a saved system prompt. You write it once and it applies to every task after it.

Score yourself, because the index won't

The AI Fluency Index deliberately reports population-level patterns and no individual score. Separately, TestingCatalog spotted an unreleased personal fluency scorecard inside Claude's settings: it reads your recent work across Chat, Cowork and Claude Code, grades 11 indicators grouped under Delegation, Description and Discernment, and returns something like 7.5 out of 11. No ship date has been announced. Don't wait for it.

Ten minutes, right now:

1. Open the last real task you handed to an AI — one whose output you actually shipped.
2. Count three briefing behaviours: did you state the outcome, fix the format, and paste one example of "good"?
3. Count three checking behaviours: did you verify a specific fact, push back on a claim, and tell it something it was missing?
4. Score the two groups separately, not out of 6. The index's whole finding is that the first number climbs while the second slips — a strong briefing score next to a weak checking score is the expected shape, not a good one.
5. Whatever your second number is, that is your real fluency ceiling.

The discernment pass

Paste this after any AI-generated diff or document, before you review it:

Before I review this, answer as a critic, not as the author:
1. Which part of this are you least confident in, and why?
2. What context did I fail to give you that would have changed the output?
3. Give me the strongest argument that this approach is wrong.
Do not defend the work. Do not rewrite it yet.

Three questions, thirty seconds, and they land on precisely the three behaviours the data says you drop on artifact work.

Why this is a career line, not a productivity tip

I manage engineers at a US company and read a lot of résumés. Everybody claims they use AI now, so the claim carries no information at all. What actually separates people in an interview is whether they can describe a time the model was confidently wrong and how they caught it. That story is Discernment, it can't be faked, and Anthropic's numbers say most people don't have one — because they stopped checking at exactly the moment the output started looking professional.

Today: run the discernment prompt above on the next thing an AI writes for you. If the answer to question 2 surprises you, your brief was the problem — and that is a five-minute fix.

Sources

• Anthropic Research — The AI Fluency Index
• TestingCatalog — Anthropic to introduce a personal AI Fluency scorecard in Claude