--- title: AI fluency framework description: The 4Ds (delegation, description, discernment, diligence), the three interaction modalities, and the description-discernment loop — consult when someone references the framework or the 4Ds, wants vocabulary for assessing their own AI skill, or asks how it relates to the Map-Test-Integrate-Compound method. category: methodology updated: 2026-09-09 --- # AI fluency framework The **AI fluency framework** is a competency model for working with AI effectively, efficiently, ethically, and safely. It was developed by Rick Dakan of Ringling College of Art and Design and Joseph Feller of University College Cork, and it is taught by Anthropic as a free online course. Its central claim is that working well with AI is a set of durable competencies rather than a collection of prompt tricks, and that those competencies survive changes in model and tool. The framework matters here for two reasons. It supplies a shared vocabulary that people increasingly arrive already holding, and it is diagnostic: naming four competencies makes it obvious which one a struggling practitioner is missing. Someone whose AI work disappoints usually has a specific gap, not a general deficit, and the four names make the gap sayable. ## The three interaction modalities The framework describes three ways a human and an AI can relate to each other during a task. In **automation**, the AI carries out a task on direct human instruction. The human knows what they want, says so, and receives it. Drafting an email, summarising a document, and producing a first pass of routine code all sit here. In **augmentation**, human and AI define and execute the work together, iteratively. Neither party holds the finished shape at the start. Essays, research, and complex problem-solving take this form, because the goal itself sharpens through the exchange. In **agency**, the human configures an AI to handle future work independently. The design object is not a task but a behaviour: what the AI should do when situations arise that the human is not present for. Tutors, assistants, and any configured system that faces users on its own belong here. The modalities are not a ranking. Each is appropriate to different work, and a competent practitioner moves between them within a single afternoon. ## Delegation **Delegation** is deciding what to hand to an AI in the first place, and it breaks into three parts. Goal and task awareness means understanding what is actually being attempted and being able to break it into parts that are suited to the AI, parts that require a human, and parts best done jointly. Platform awareness means knowing honestly what current tools can and cannot do, including the constraints imposed by budget, procurement, and regulation. Task delegation is the assignment itself: matching each part to the party best suited to it, and choosing which modality fits. Delegation is where most failures originate, because a badly chosen task cannot be rescued by good prompting. It is also the competency most improved by simply decomposing work before starting, which is why structured workflow decomposition pays for itself. ## Description **Description** is communicating the work to the AI, and it also has three parts, distinguished by what is being described. Product description defines the output: what it must contain, how long it should be, what shape it takes. Process description governs how the work proceeds: the steps to follow, the order, the back-and-forth, and how a large request is broken into manageable pieces. Performance description defines behaviour rather than output, which matters when configuring something that will run later without supervision. It answers what the AI should do in situations the person is not present for. The distinction between product and performance is the one people miss. A prompt written for today's example describes a product. A prompt written for a persistent assistant or a scheduled run must describe performance, because it will meet inputs its author never saw. Writing the first when the second is needed is a common and expensive error. Six techniques carry most of the practical weight: giving context, showing examples, defining constraints, decomposing complex tasks, asking for reasoning before conclusions, and specifying role or tone. The last of these is frequently overrated. A role line aims the model at a body of knowledge and sets tone, but it confers no capability and never substitutes for the reference material the task actually requires. ## Discernment **Discernment** is judging what comes back, and it mirrors description across the same three dimensions. Product discernment evaluates the output itself: whether it is accurate, whether it is relevant, and where it is weak. Process discernment evaluates the collaboration rather than its result, asking whether the working relationship is productive or whether the person is fighting the tool. Performance discernment evaluates a configured system in use, gathering evidence about how it behaves for the people it serves and refining accordingly. Process discernment is the least practised of the three. A workflow can produce acceptable output while costing more attention than it saves, and nothing in the output reveals this. The signal lives in the number of rounds required, the frequency of restarts, and whether the person quietly avoids running it. ## The description-discernment loop Description and discernment are not sequential stages but a cycle: describe, receive, judge, describe again. The framework treats this loop as the core operating habit and everything else as scaffolding around it. The practical consequence is that a single AI response is never evidence of a working process. One run is a demonstration; the second run, made after judging the first and changing the request accordingly, is the first real test. Practitioners who treat AI as a vending machine never enter the loop, and consequently never improve their instructions. The loop also reframes disappointment productively. When output is wrong, the useful question is not what the AI failed to do but which assumption went uncommunicated, and whether the gap was in the product, the process, or the performance description. ## Diligence **Diligence** is responsibility for AI-assisted work, and it is the competency most often absent from practical AI training because it produces no visible output. Creation diligence covers responsible use during the work itself: which tool is appropriate for which data, awareness of bias and legal constraint, and consideration of who is affected. Transparency diligence covers telling people what the AI did, calibrated to the expectations of the audience and the norms of the field. Deployment diligence is verification and ownership: a named person checks the work, vouches for it, and carries the risk if it is wrong. The three are easy to state and easy to skip. A workflow that passes every quality check can still fail on diligence, because the failure is not in the output but in what was never asked. See: ai-diligence. ## Where the four competencies sit in Map, Test, Integrate, Compound The competencies and the four-phase method answer different questions. The method is a sequence in time; the competencies are capabilities exercised throughout. Mapping them one to one is a mistake, but they do have centres of gravity. Mapping a workflow is delegation work, plus the first diligence question about what data the work touches. Testing is description and discernment together, running as a loop rather than a stage. Integration is deployment diligence and performance discernment, since the questions that matter are whether the thing behaves correctly in daily use and who owns it when it does not. Compounding has no counterpart in the framework at all. That last point is worth stating plainly. The description-discernment loop operates within a working session: you judge an output and improve the request in front of you. Compounding asks a different question, whether anything captured in this run makes the next run better, and it requires a mechanism that outlives the conversation. A practitioner can be fluent by the framework's definition and still have no workflow that improves with use. See: compound-engineering. ## The automation naming collision The word automation means close to opposite things in the two vocabularies, and this causes real confusion. In the AI fluency framework, automation is the modality where a human instructs and the AI executes immediately, such as asking for one summary. In the four modes of AI work, the third mode long carried the same name but meant a workflow running on a schedule or trigger without the human present. Someone holding both would picture the wrong thing on hearing either. The mode was renamed clockwork on 2026-09-09 to resolve this. The artifact built in that mode is still called an automation, which is unambiguous because it names a thing rather than a mode. See: ai-work-modes. ## Related pages - See: ai-work-modes - The four modes of AI work, and how they line up against the three modalities - See: ai-workflow-redesign - The methodology that schedules these competencies across phases - See: ai-diligence - The fourth competency in operational detail - See: ai-output-verification - Discernment in practice, including how to judge a collaboration rather than an output - See: compound-engineering - The mechanism the framework does not address