The AI course Anthropic makes its own staff take
The short version: on 20 August 2026 Anthropic put the AI training it gives every new employee in their first week online at academy.claude.com. It is free, and you do not need an account to take it. The main course is 14 lessons, about four hours. What it teaches is not prompt tricks but four habits, called the 4D framework: Delegation, Description, Discernment and Diligence. And buried in the announcement is an admission worth reading twice, that the specific prompt phrases people have spent two years memorising are becoming unnecessary as models improve.
What this actually is, in plain terms
Anthropic, the company that makes Claude, runs an internal training programme. Every person who joins is taught, during onboarding, a framework for working with AI: what to hand to it, how to ask, how to check the answer, and what you remain responsible for. In their words, the journey to AI fluency begins on an employee’s first day. After onboarding there are what they call ever-boarding programmes, which is their term for training that never really ends, because the tools change too fast for a one-off induction to hold.
That programme is now a public website. Here is the entire process of using it, with nothing left out:
- Go to
academy.claude.com. - Pick a course or a tutorial.
- Press Start course.
That is genuinely all. There is no sign-up wall, no card, no trial. The page offers you a separate “Sign in to save progress” link, and that is the only thing an account buys: it remembers where you stopped, and it gives you a completion badge at the end. Every lesson is readable without it.
The main course is AI Fluency: Framework & Foundations — 14 lessons, one quiz, about four hours. Alongside it sit a few dozen much shorter tutorials on single questions, most of them five to twenty minutes: why do AI models hallucinate (5 min), what is sycophancy in AI models (6 min), can you trust what AI tells you (5 min), what does AI know about me (5 min).
The line that quietly expires two years of prompt advice
The announcement contains a paragraph that reads like housekeeping and is actually the most useful thing in it. Anthropic explain why they stopped teaching specific prompt behaviours, and they give an example.
You used to have to tell Claude who the work was for. Their example is telling it that a task was intended for colleagues working in legal, because otherwise the register came out wrong. With the newest models, they write, Claude will simply ask you for that information if it is needed for a high quality output. The trick did not stop working. It stopped being necessary, which is worse for anyone selling it.
Their conclusion is the interesting part: their education team has shifted away from emphasising specific behaviours for AI fluency and towards cultivating broader, more durable mindsets. In other words, the company with the most commercial reason to sell you a list of magic phrases has decided the list is a dead end. Which is roughly what we argued in the memo on secret ChatGPT codes, except this time it is coming from the people who build the model.
The four habits, in plain English
The framework is four words, all starting with D, which is why it is called 4D. The names are academic and the ideas underneath them are not. Here is each one, what it actually means, and what it looks like on a Tuesday afternoon.
Delegation — decide what you hand over
Before you type anything, decide which parts of the job are the AI’s and which parts stay yours. The course splits this into three questions: what am I actually trying to achieve, what is this particular tool good and bad at, and therefore which pieces go where.
Anthropic’s own example is a memo: you draft the sensitive paragraphs yourself and let AI assemble the summary slides. Or you let it do exploratory analysis on a dataset and you do the final checks. The habit is noticing that this is a decision at all, rather than defaulting to either extreme.
There is a second reason they push this one, and it is not about output quality. Their materials explicitly encourage you to keep practising the skills that matter to you, to avoid what they call skill atrophy. If writing is the thing you are good at, handing all of your writing away has a cost that does not show up this week.
Description — say what you actually want
This is prompting, but framed as three separate things rather than a phrase to copy. Product: what should come out — length, format, who reads it. Process: how it should get there — think first, ask me questions before starting, show your working. Performance: how it should behave — blunt rather than agreeable, a critic rather than a cheerleader.
The difference in practice is the gap between “summarise this” and “give me 150 words for a client who has not read the report, no jargon, and ask me anything you are missing before you start”. Nothing in the second one is a trick. It is just the information a competent human would need, and it does not expire when the model gets better.
Discernment — judge what comes back
Confident, tidy and wrong is the failure mode that costs people real money, and it is the one the Academy spends the most time on. Discernment is judging three things: the output itself, the route it took to get there, and its behaviour — whether it is agreeing with you because you are right or because you sounded certain.
The mechanic they teach is the description-discernment loop: look at what came back, work out what was wrong with your description, describe it again, look again. Most people run one round and stop.
Diligence — own the result, and say where AI was
The last one is about everything around the conversation. Was this built responsibly, are you transparent about how it was made, and do you stand behind what you send? The Academy calls these creation, transparency and deployment diligence.
The concrete deliverable here is the AI diligence statement: a short, specific note saying how AI was used, published alongside the work. There is a 15-minute tutorial on writing one. Anthropic put one on the fluency course itself, disclosing that Claude 3.7 helped with structural development, exercise design and drafting, while the named human authors made every final decision.
The whole trick is specificity. Here is the shape, and you can lift it:
Claude was used to draft section 3 and to check the arithmetic in the table. The figures come from the Q2 export and were verified against it by hand. The recommendations and the final wording are mine.
Compare that with “this document was AI-assisted”, which tells a reader nothing they can act on.
What to take first, and what to skip
Four hours is a real ask, and you do not need all of it. If you have one evening, this order gets you the most per minute:
- The 4D framework lesson, inside Framework & Foundations. It is the spine; everything else hangs off it.
- The 4 Properties of AI — 7 minutes, and it is the clearest short thing on the site about where these tools break.
- Why do AI models hallucinate (5 min) and what is sycophancy in AI models (6 min). Two failure modes, eleven minutes, and you will start spotting both.
- Writing an AI diligence statement (15 min), if you send work to anyone who would care.
Skip, unless they are your job: the five product tracks, the enterprise rollout material, and the train-the-trainer courses. There are also tailored editions of the fluency course — for builders, educators, pK-12 educators, nonprofits, small businesses and students — which cover the same ground with examples from your world. Pick one edition, not several.
Two tutorials deserve a mention for anyone watching what they spend: tokens: why some inputs cost more than others (15 min) and how context affects Claude’s performance and cost (20 min). If that is your problem, our memo on cutting Claude Code token usage is the practical companion to both.
Who actually wrote this
Worth knowing, because it changes how much weight to give it. The 4D framework is not Anthropic’s invention. It was developed in 2023 and 2024 by Rick Dakan, a professor at Ringling College of Art and Design, and Joseph Feller, at University College Cork, from their research into how AI tools were changing creative and business work. Anthropic saw it, partnered with them, and built the course around it. The work was supported in part by the Higher Education Authority in Ireland.
So the four habits were taught to undergraduates and postgraduates at two universities before they were ever taught to Anthropic staff. That is a better provenance than most corporate training, and Anthropic’s own announcement barely mentions it.
One broken link, and the fix
The announcement tells you to install the Claude Academy Skill so Claude can recommend courses based on how you work, and links it. That link was dead when this memo was written: it points at a folder called claude-academy-guide in Anthropic’s public skills repository, and no such folder exists there.
The skill is real, though. The folder is named academy-guide. This is the working address:
https://github.com/anthropics/skills/tree/main/skills/academy-guide
Checked 22 August 2026. If Anthropic fix their link, both will lead to the same place.
What it costs
Nothing, and there is no upsell inside the lessons. You can also reach the Academy from the “Learn more” tab in the Claude profile menu if you already have an account. A Claude subscription is not required to take a single course on the site, including the ones that teach Claude.
The only honest cost is the four hours, and the reason to spend them is the thing the announcement says out loud: the specific tricks keep expiring, and the habits do not.
Questions people actually ask
Is Claude Academy free?
Yes, and you do not need an account to read or watch any of it. Open academy.claude.com, pick a course and press Start course. Signing in is offered separately, and it only buys you two things: it remembers which lessons you have finished, and it gives you a completion badge at the end of a course. Nothing is paywalled and there is no card involved at any point.
What is the 4D AI fluency framework?
Four habits that make up working well with AI: Delegation (deciding what to hand over and what to keep), Description (telling it clearly what you want), Discernment (judging what comes back, and how hard to check it), and Diligence (taking responsibility for the result and being honest about where AI was involved). It is the framework Anthropic teaches every new employee during onboarding, and the free Framework and Foundations course is 14 lessons and about four hours.
Who created the 4D framework?
Not Anthropic. It was developed in 2023 and 2024 by two academics: Rick Dakan of Ringling College of Art and Design and Joseph Feller of University College Cork, out of their research into how AI tools were changing creative and business work. Anthropic partnered with them to turn it into a course. The work was supported in part by the Higher Education Authority in Ireland through the National Forum for the Enhancement of Teaching and Learning.
Are prompt tricks like act as an expert obsolete now?
The specific ones are fading, and Anthropic said so in the announcement itself. Their example: you used to have to tell Claude that a task was intended for colleagues working in legal, or the tone came out wrong. With the newest models it simply asks you for that information when it needs it. The advice that survives is not a phrase to paste, it is a habit: be clear about what you want, and check what you get.
How long does the main AI fluency course take?
AI Fluency: Framework and Foundations is 14 lessons plus one quiz, and the site estimates about four hours in total. You do not have to do it in one sitting, and you do not have to do all of it. The 4D framework lesson alone gives you the structure; the deep dives on generative AI, prompting and the description-discernment loop are where the four hours go.
What is an AI diligence statement?
A short, specific note saying how AI was used in producing something, published with the work rather than buried. Anthropic uses one on the fluency course itself, disclosing that Claude 3.7 helped with structure, exercise design and drafting while named human authors made every final decision. There is a 15-minute tutorial on the Academy on how to write one. The point is that it is specific: saying a document was AI-assisted tells a reader nothing.
Is Claude Academy just marketing for Claude?
Part of it is, and it is worth knowing which part. The site has five product tracks (Claude.ai, Claude Cowork, Claude Code, Claude Tag and Claude Platform) that teach you their own software. The AI Fluency collection is different: it is deliberately product-agnostic, the framework came from outside the company, and almost none of it depends on using Claude. If you use ChatGPT or Gemini, the fluency half still applies to you unchanged.
Where is the Claude Academy skill the announcement mentions?
The link in Anthropic's own announcement points at github.com/anthropics/skills/tree/main/skills/claude-academy-guide, which returned a 404 when this memo was written on 22 August 2026. The skill does exist, at a slightly different path: the folder is named academy-guide, not claude-academy-guide. Install it and Claude will recommend courses and learning paths based on how you actually work.
Sources
- Anthropic — Anthropic's approach to teaching and learning AI (the announcement, 20 August 2026)
- Claude Academy — the site itself, free and open without an account
- Claude Academy — AI Fluency: Framework & Foundations (14 lessons, about four hours, and the course's own AI diligence statement)
- Claude Academy — the AI Fluency collection, including the short tutorials on hallucination, sycophancy and bias
- Anthropic — the academy-guide skill (the working link; the announcement's own link 404s)
Want this built for you?
We write these memos because we build this stuff every day. If you want it working in your business instead of sitting on your reading list, that is literally our job.