The 4D AI Fluency Framework Explained
Working with Claude well is a skill, not a switch you flip once.
//
Search across all documentation pages
Working with Claude well is a skill, not a switch you flip once.
Anthropic Academy's course "AI Fluency: Framework and Foundations," developed with academic partners, organizes that skill into four interlocking practices known as the 4D Framework.
The four practices are Delegation, Description, Discernment, and Diligence.
Each one answers a different question a person asks while collaborating with Claude, from "should I even hand this off" through "did I actually get what I needed."
This page walks through what each of the four means, how they relate to one another, and why treating them as a single loop rather than four separate skills changes how you use Claude day to day.
The 4D Framework grew out of Anthropic Academy's work training students, educators, nonprofits, small businesses, and builders to use Claude effectively rather than either avoiding it or leaning on it uncritically.
The course frames fluency with AI the same way fluency with any collaborative tool works: it is less about knowing every feature and more about knowing when and how to use it well.
Delegation is the first practice.
It is the decision of what to hand to Claude versus what a person should keep doing themselves.
Description is the second practice.
It is the craft of telling Claude what you actually need, with enough context and constraints that it can act on your intent rather than guess at it.
Discernment is the third practice.
It is the habit of reading what Claude produces with a critical eye, catching errors, missed nuance, or bias before that output goes anywhere important.
Diligence is the fourth practice.
It is the ongoing responsibility to verify, iterate, and use the whole process safely and ethically, rather than treating one good result as proof the job is done.
A simple analogy: think of the four D's as stages of handing work to a new collaborator.
You decide what to give them, you brief them clearly, you check their work before you rely on it, and you stay accountable for the outcome even after they hand it back.
Claude is not a person, but the same discipline applies, and skipping any one stage tends to produce the same kinds of problems it would with a human collaborator.
The four D's are not four independent skills you learn in isolation.
They form a loop, and each one shapes the ones around it.
A Delegation decision determines how much Description a task will need.
A well-scoped, low-stakes task delegated to Claude might need only a sentence of instruction.
A high-stakes or ambiguous task, even one worth delegating, usually needs a much fuller Description: context, constraints, examples, and a clear statement of the actual goal.
Description quality, in turn, shapes how much Discernment work is required afterward.
A vague prompt tends to produce output that looks plausible but drifts from what you actually wanted, which means more careful review is needed to catch the drift.
A precise prompt with the right context narrows the range of plausible outputs, which makes Discernment faster and more reliable, though it never eliminates the need for it.
Discernment findings feed back into both Delegation and Description.
If review repeatedly turns up the same kind of error, that is a signal either that the task was not actually a good fit for delegation, or that the instructions given were missing a constraint that should have been stated up front.
Diligence sits over the whole loop rather than at one point in it.
It is what makes the loop actually a loop instead of a one-time pass: verifying results, iterating on instructions that did not land, and staying attentive to safety and ethical considerations across repeated use, not just the first attempt.
Delegation -> Description -> [Claude works] -> Discernment -> (iterate or ship)
^ |
'------------------ Diligence -------------------'This loop runs at very different speeds depending on the task.
For a quick email draft, all four D's might happen in under a minute, almost unconsciously.
For a substantial research synthesis or a project brief, each stage deserves deliberate attention, and the Diligence stage may span several rounds of iteration.
The framework becomes more interesting once you notice where each practice tends to fail in isolation.
Delegation fails in two opposite directions.
Under-delegation wastes Claude's value by keeping it to trivial tasks when it could reliably help with first-pass drafting, research synthesis, or repetitive analysis.
Over-delegation hands off final judgment calls, sensitive decisions, or anything requiring accountability that genuinely cannot be delegated, then treats Claude's output as if it were the final word.
Description fails most often through omission rather than error.
People state an instruction but leave out the goal behind it, the constraints that matter, or the context that would let Claude distinguish a good answer from a merely plausible one.
This is also where context-window awareness becomes practical rather than theoretical: deciding what background information is worth including in a prompt, and what would only dilute it, is itself a Description skill.
Discernment fails when it is skipped for output that looks fluent.
Claude's responses are typically well-written and confident in tone regardless of whether the underlying content is fully correct, which makes fluent-sounding errors easy to miss without deliberate review.
Diligence fails when a single successful result is treated as proof that a process is trustworthy going forward, rather than as one data point that still needs verification the next time conditions change.
| Practice | Strength | Weakness | Best Fit |
|---|---|---|---|
| Delegation | Frees time for higher-judgment work | Easy to over- or under-apply | Recurring tasks with a clear boundary between routine and judgment work |
| Description | Directly raises output quality | Takes upfront time and thought | Any task where the goal or constraints are not obvious from a one-line request |
| Discernment | Catches errors before they matter | Can be skipped when output looks fluent | Any output that will inform a decision or be shown to someone else |
| Diligence | Builds a trustworthy long-term process | Requires sustained attention, not a one-time check | Repeated or high-stakes use of Claude over time |
Anthropic Academy offers variants of the AI Fluency course tailored to students, educators, nonprofits, small businesses, and builders, because the specifics of good Delegation and Description differ by context even though the same four practices apply.
A student deciding what to delegate on a research paper faces different stakes than a nonprofit delegating grant-writing support, but both are running the same loop.
A four-part discipline, Delegation, Description, Discernment, and Diligence, for deciding what to hand to Claude, how to instruct it, how to judge its output, and how to keep the process trustworthy over time.
Deciding which tasks suit handing off to Claude, such as repetitive drafting, research synthesis, or first-pass analysis, versus which tasks need to stay human-led, such as final judgment calls or sensitive decisions.
Giving Claude sufficient context, explicit constraints, and the actual goal behind a request, rather than a bare instruction, so it can act effectively on what you actually need.
Critically evaluating what Claude produces: catching factual errors, missed nuance, or bias before that output is relied on or passed along to someone else.
The ongoing responsibility to verify results, iterate on instructions that did not work, and use the whole process safely and ethically across repeated use, not just once.
A vague prompt tends to produce output that looks reasonable but drifts from what was actually wanted, which means the review stage has to work harder to catch that drift; a precise prompt narrows the range of likely output and makes review faster.
Yes. Handing off final judgment calls, sensitive decisions, or anything requiring accountability that cannot be delegated, and then treating Claude's output as the final word without review, is a common failure mode the framework is designed to prevent.
Yes. Keeping Claude to trivial tasks out of caution, when it could reliably help with first-pass drafting, research synthesis, or repetitive analysis, wastes the tool's value and is treated as its own kind of Delegation failure.
Claude's output is typically fluent and confident in tone regardless of whether the underlying content is fully correct, so errors can be easy to miss without deliberately reading for them rather than just skimming.
The same four practices apply to every task, but the depth needed scales with the stakes: a quick low-stakes request may run through all four D's almost instantly, while a substantial project deserves deliberate attention at each stage.
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.
Reviewed by Chris St. John·Last updated Jul 16, 2026