I have never taken a course on the thing I spend most of my working day with. Everything I know about working with a model came out of evenings where something did not behave the way I expected. That teaches you plenty, slowly, and mostly it teaches you your own habits.
On 20 August Anthropic announced Claude Academy. Free, no paywall, courses on their own products and on AI in general. I opened the catalogue to see what I had missed. I got stuck in the research report sitting underneath the curriculum, because it contains a measurement that does not fit the announcement.
What is actually there
The catalogue splits in two. Product tracks for Claude.ai, Cowork, Claude Code, Claude Tag and the platform with the API and MCP. Alongside that, a collection about AI fluency itself, which deliberately is not about Claude.
On 21 August I counted what you can reach from the homepage, the five product hubs and the fluency collection: nineteen courses and thirty-two tutorials. A course is a run of lessons with a quiz at the end, a tutorial is a standalone piece of a few minutes.
| Course | Size | What it covers |
|---|---|---|
| AI Fluency: Framework & Foundations | 14 lessons, 1 quiz, 4 hr | The 4D framework: delegation, description, discernment, diligence |
| AI Capabilities and Limitations | 13 lessons, 1 quiz, 3.5 hr | What a language model does and does not do, down to context limits |
| Claude Code 101 | 12 lessons, 1 quiz, 1 hr | Building and debugging from the terminal |
The material reads without an account. I fetched the pages while logged out and got the full text back. Tracking progress and collecting badges does need a profile.
Two things in the accompanying blog post struck me more than the catalogue itself. Anthropic says the material matches what new employees get on day one, including the 4D framework and a programme they call “ever-boarding”, an onboarding that never finishes. And the education team has walked away from teaching specific tricks. Telling Claude who your audience is used to matter; now the model asks you when it needs to know. What survives are mindsets. Two of them appear in the material word for word: today’s AI is the worst AI you will ever use, and verify in proportion to the stakes.
The measurement under the curriculum
Underneath the curriculum sits the AI Fluency Index, first published on 23 February 2026.
The setup decides what the numbers are worth. The 4D framework comes from professors Rick Dakan and Joseph Feller, developed with Anthropic, and describes 24 behaviours that add up to safe and effective collaboration with AI. Eleven of those are readable from a conversation with Claude. The other thirteen happen outside the chat window, things like being honest about the role AI played in your work, so they went unmeasured. Anthropic calls those thirteen arguably the most consequential ones.
For the first round, 9,830 Claude.ai conversations were analysed with a privacy-preserving tool, drawn from a seven-day window in January 2026, restricted to conversations with several back-and-forths. The result is a baseline, not a trend.
What came out:
- 85.7 percent of conversations show iteration and refinement. Those conversations average 2.67 additional fluency behaviours against 1.33 without, roughly double.
- Within them, people are 5.6 times more likely to question the model’s reasoning and 4 times more likely to spot missing context.
- 12.3 percent of conversations produce an artifact, such as code, a document or a working tool.
The number that suits nobody selling these products sits in those same artifact conversations. People direct the model harder there. They state their goal more often (+14.7 percentage points), specify a format more often (+14.5) and supply examples more often (+13.4). And they check less. Spotting missing context drops by 5.2 percentage points, fact checking by 3.7, questioning the reasoning by 3.1.
The more finished the result looks, the less anyone doubts it. Anthropic offers three possible explanations in the report, one of them being that verification moves outside the conversation: you run the code, you test the app, you show a draft to a colleague. That is fair. It does not explain why fact checking in particular falls on output that contains facts.
What the second round adds
The index was later extended to Claude Code and Cowork, over 50,000 conversations in total. That is where the finding I keep coming back to lives, especially if you are bringing a team along.
Each product has a different entry behaviour. In chat it is iteration and refinement. In Code and Cowork it is stating your goal clearly before the model starts working. Whoever learns that first sees every other skill lift with it. It follows from which behaviour correlates most strongly with all the others.
The larger point is how two kinds of skill develop. Description grows on its own. People who have used Claude longer supply examples more often, shape tone more deliberately, set up projects more often. Discernment does not grow on its own. It does not track tenure and it does not come along with product knowledge. It has to be taught explicitly, over and over, and it is precisely the skill that drops the moment the output looks finished.
Anthropic adds that the tasks which used to go to early-career employees are being automated. Those tasks were where you learned what good work looks like. Automate them away and you have to teach it instead.
Where the announcement is wrong
The blog post says you can have Claude recommend courses through a skill, and links to anthropics/skills/tree/main/skills/claude-academy-guide. That page returns a 404. The skill does exist, but it is called academy-guide, one folder over in the same repository. Checked on 21 August 2026.
The contents are better than the broken link suggests. The skill instructs the model to check for a matching course before it finishes answering a question about using Claude. It also says explicitly that it may only recommend on a strong match and must never invent Academy content. How I read instruction files like this one I wrote up earlier when arguing for telling the model less.
The caveat
The vendor writes the exam. The company that builds the model also defines what good use is here, and measures it in its own logs. The behavioural framework comes from outside academics, which helps, but the sample consists of conversations on Claude.ai. A score for skilled AI use that can only be measured inside one product also measures how well you operate that product.
The bill for mistakes lands on you. “Verify in proportion to the stakes” is a good rule and also a transfer. The company supplies the answer, the user carries the consequences of an answer that is wrong. Their own research shows that this verification drops when the result looks polished, so the rule is demanded hardest at the moment people apply it least.
What drains away is your judgement, not your work. Description you pick up by doing. Discernment you do not. And the small jobs where you used to learn what good work looks like are the first to move to the model. The report says so itself. A free course answers the first half of that well. It does not answer the second.
What I am doing about it
I am doing two things about it and skipping a third.
I am going to take the capabilities and limitations course, 3.5 hours, spread over a few evenings. Not to operate Claude better, but because the report suggests my weak side is probably discernment rather than direction. There is no reason to think I differ from those 9,830 conversations.
I also have an explicit verification step in my editorial chain after any piece that looks finished, which I had been treating as hygiene. Since reading this report I treat it as the step that matters most. It sits next to the guardrails that changed under me without an announcement.
What I am not doing is installing the skill that recommends courses. I can read a catalogue myself.
Frequently asked questions
Does Claude Academy cost anything?
No. The courses and tutorials are free with no paywall. The material also reads without an account; only progress tracking and badges need a profile.
Is this only useful if you use Claude?
No. The five product tracks are about Claude, but the fluency collection is deliberately product and model agnostic. Courses on what a language model can and cannot do, on hallucination and on sycophancy travel fine elsewhere.
What exactly is the 4D framework?
Four groups of behaviours: delegation (what you hand over and what you keep), description (how you explain what you want), discernment (how you evaluate what comes back) and diligence (how you handle sources, disclosure and consequences). It was developed by Rick Dakan and Joseph Feller with Anthropic and contains 24 concrete behaviours.
Why does the artifact finding matter?
Because it is about your work rather than the model’s. As soon as something looks finished, people check it less. On a document full of facts, that is exactly the wrong order.
Sources
- Claude, announcement on X — 20 August 2026
- Anthropic’s approach to teaching and learning AI — claude.com, 20 August 2026
- Claude Academy — catalogue consulted 21 August 2026
- Anthropic Education Report: The AI Fluency Index — originally published 23 February 2026
- Getting good at Claude: A research-backed curriculum — consulted 21 August 2026
- anthropics/skills, academy-guide folder — consulted 21 August 2026
Checked on 21 August 2026. I counted the courses and tutorials myself from the homepage, the five product hubs and the fluency collection, so there may be material reachable only through search or a signed-in profile. I could not watch the 34 second video attached to the announcement, so everything above comes from written sources. The blog post and the repository disagree on the path to the skill; I went with the repository, because that one can be checked.