uygnoey Designing AI-fluency training
Derived from Anthropic's account of how it approaches teaching and learning AI, published alongside the launch of Claude Academy.
The organizing claim: AI instruction should increase agency and empower learners to expand their capabilities. Everything below is a consequence of taking that seriously instead of teaching the interface.
Instructions
1. Start from the learner's problem, not the product surface
A feature tour answers "what can this do?" Agency-first material answers "what were you trying to get done?"
- Open each unit with a real task the learner already owns. The AI capability is introduced as a way through that task, not as the subject.
- Make delegation an explicit decision the learner practices, not a default. Ask directly: which parts of this task should stay yours, and why?
- Judge a unit by whether the learner can now do something they could not do before — not by how many features were covered.
Anti-pattern: a module named after a product surface. If the title is a feature name, the unit is probably a tour.
2. Teach mindsets that outlive the model
Feature-level instruction decays with every release. Framings do not.
Two carried by the source:
- "Today's AI is the worst AI you'll ever use." Treat a current limitation as a snapshot, and re-test assumptions rather than encoding them as rules.
- "Verify in proportion to the stakes." Not "always check everything" and not "trust it" — a calibration the learner performs per task.
When you catch yourself writing a numbered click-path, ask what judgment the learner would need if the interface changed. Teach that instead. See examples/mindsets-vs-instructions.md for rewrites.
3. Cover safe use holistically
Safe use is not confined to the interaction itself. The curriculum has to reach three areas that prompt-craft training normally skips:
- Task selection — deciding what is appropriate to hand over at all.
- Skill preservation — continuing to practice skills that matter to you, so they do not atrophy through delegation.
- Disclosure — explicitly stating how AI was used in producing a document, analysis, or piece of media before sharing it with others.
Details and how to teach each: references/holistic-safe-use.md.
4. Make the learning active
Reading about AI use does not produce AI use. Every unit should carry:
- a practice exercise on the learner's own material, not a toy prompt;
- a reflection prompt that forces a judgment call rather than a recall;
- room to experiment, so the learner arrives at an approach fitted to their work instead of copying the instructor's.
Use templates/course-outline.md as the scaffold — it will not let you write a unit without all three.
5. Aim at fluency, then let it amplify
The final goal is not tool proficiency but the point at which the learner can use AI to learn anything else. Once fluency is established, teach people to treat AI as a learning partner in whatever domain they need next — that is where the curriculum's leverage actually comes from.
6. Structure the program as onboarding plus continuing education
One-time training goes stale by definition. Anthropic's own shape:
- Onboarding — a fluency framework taught to every new employee (internally, the 4D AI Fluency Framework).
- "Ever-boarding" — ongoing programs that keep covering current capabilities, current limitations, and best practices for working with agents.
The continuing track is where model-specific content belongs, precisely because it is the part you expect to rewrite.
The five design principles in full, each with what it rules in and out: references/design-principles.md.
Examples
Rewriting a stale module
A team has a module titled "Using the file-upload feature." Every screenshot is one release out of date and the module is skipped in practice.
Rewritten under this skill: the unit becomes "Getting an answer out of a document you did not write." It opens with a contract or report the learner actually has to review. The upload step appears in one line. The substance is the judgment: what parts of this review must you do yourself to be accountable for the output, and how hard should you verify each claim given what it is attached to? The exercise uses the learner's own document; the reflection prompt asks them to name one thing they will keep doing manually and say why.
A safe-use unit that is not about prompting
Instead of "how to write a good prompt," the unit asks the learner to take three tasks from their week and sort them: delegate fully, delegate with review, keep. They then justify one item in the "keep" column purely on skill-preservation grounds — a skill they want to stay sharp at — and draft the one-line disclosure they would attach to the delegated output before sending it on.
Choosing where content lives
A new model ships with a longer context window. That fact goes in the ever-boarding track, not in onboarding: onboarding teaches that context is finite and that you decide what deserves to be in it, which stays true regardless of the number.
Source
Anthropic's approach to teaching and learning AI (published 2026-08-20). Claude Academy is at academy.claude.com.