# 0.4.2 The four Ds (/how-ai-works/ai-fluency/the-four-ds) --- type: Document title: "0.4.2 The four Ds" description: "Delegation, Description, Discernment and Diligence, with their twelve parts in the framework's terms, and when each is done in a chat and when the AI works alone." status: stable order: 14.2 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-11-03T21:38:22Z chapter: "0.4" part: "0" expert_status: null # as the Part 0 overview: Part 0 is optional on every route, after the skip quiz concepts: [ "0.4.2" ] last_verified: 2026-10-16 sources: - id: anthropic-ai-fluency resource: https://www-cdn.anthropic.com/4396730ed190e691a3712cf2fd6bfe35509deca2.pdf title: "AI Fluency: Key Terminology Cheat Sheet, the framework's two-page terminology sheet (Rick Dakan, Joseph Feller and Anthropic, CC BY-NC-SA 4.0, verified 2026-10-16)" generated: at: 2026-11-03T21:38:22Z by: process:panaversity trust_tier: unverified build_id: sha256:e1fbcb7d9a8f8b9b7cb96875022325c02c9ba4be94cef13b301cbd39a9edad85 dirty: true ksor_version: 0.0.60 --- **In everyday life.** When you look after your child yourself, you decide, explain, watch and take responsibility for things as they happen. When a babysitter comes for the evening, you do the same things at other times. The babysitter has never met your child. So before you leave, you write down what the babysitter may do alone, and the bedtime routine. The next morning, you ask what happened and see for yourself. You are the parent all evening. **What you will learn.** When you do each of the four Ds, in a chat and when the AI works alone. Who needs to know that AI was involved. **Why it matters.** When the AI works alone, nobody acts during the run. So each D is done before the run or after it, or it is not done at all. ## Delegation Delegation means deciding which work people do, which work the AI does, and how to split the tasks between them.[^anthropic-ai-fluency] You understand your goal, you know what different AI systems can and cannot do, and you decide task by task who does what. ### When is Delegation done? Before the work starts. When the AI works alone, that means in writing, before the first run. So a written plan says what the AI does alone, and what waits for a person, such as any correction to the register. In a chat, you can also change it at any turn, as you did at 2:05. Delegation is more than handing work over. You delegate in every mode, and part of it is deciding what stays with a person. ## Description Description means communicating effectively with AI.[^anthropic-ai-fluency] You say three things: what you want back, how the AI should approach it, and how it should behave. ### Why must Description be written down before the run? Because a run has only what is in front of it when it starts, and it has no next turn. In a chat, you add what is missing as you go. When the AI works alone, the description is finished before the first run. It is kept where every run reads it, such as a project's instructions, which 0.2 showed. ### What happens to the AI's behavior when nobody is watching? It becomes written behavior. In a chat, you tell the AI how to behave, and you can change that at any turn. When the AI works alone, you are not there to change it, and the people who read its work may never type to it. So its behavior is decided in advance and written down, such as what it does when it is unsure. At Brightline, when the statement does not explain a difference, the AI writes "the statement does not show" instead of a guess. ## Discernment Discernment means judging, with care, what the AI produced, how it got there and how it behaved.[^anthropic-ai-fluency] ### When is Discernment done, if nobody watched the work? After each run, at a set time. In a chat, you judge each answer as it arrives, and you saw every step that led to it. When the AI works alone, nobody watched. The run may check its own work, but its check does not replace yours. So you judge afterward, from the output, the run's record and the source. The run's record shows what the run did, such as which files it read and what it wrote. The source holds the facts: the statement, the register and the invoices. ### How can you judge how the AI arrived at its output, if you did not watch? From the run's record, and across runs. The record shows what the run read and wrote, which is evidence of how the output was made. A mistake that comes back every month shows a problem in how the output is made. The record also lets you check whether the run behaved as its written instructions said. Chapter 6 teaches how to read a run's record. ## Diligence Diligence means using AI responsibly and ethically.[^anthropic-ai-fluency] You choose with care how you use AI and what you put in. You are honest about AI's part in your work. And you take responsibility for what you use or share. ### Why does Diligence run through the whole job? Because it has three jobs that come at different times, and when the AI works alone, nobody is there in between. First, before the first run, the plan says how the AI is used and what may go in. Next, the honest word about AI's part travels with the output. Last, after each run, someone signs for the work, standing behind it before it is used. Discernment judges the work, and signing takes responsibility for it. In a chat, you do all three in one sitting. ### Who needs to know that AI was involved? Everyone who relies on the work, including people who never typed to the AI. In a chat, you tell the person you send the work to, as you told Dave at 2:20. When the AI works alone, its work can reach people with nobody there to tell them. At Brightline, Dave signs the reconciliation, and auditors, who check a company's accounts, may rely on it later. So the plan says who is told, and the table itself says that AI made it. Chapter 10 teaches when and how to tell people. ## The same four Ds, moved in time ### What changes when the AI works alone? Only when you do each D. Delegation and Description move before the first run, in writing. Discernment moves after each run, to a set time. Diligence runs through the whole job. The Preface's operating rule says: "Humans define intent. AI Workers execute. Humans verify outcomes." Delegation and Description define the intent, and Discernment verifies the outcome. ![The same four Ds, moved in time. Two lanes share one timeline. In the top lane, "In a chat", a person and the AI pass arrows both ways at every turn, with all four Ds at each turn. In the bottom lane, "When the AI works alone", a page before the run holds "Delegation" and "Description", and one arrow runs from it into "the run". No person stands beside the run, and no arrow runs from a person into it. One arrow runs out to "the output and the run's record". After the run, "Discernment, at a set time", where a person reads, is fed by three arrows: from "the output", "the run's record" and "the source". Under the whole lane runs a band, "Diligence: what goes in, who is told, who signs". A banner reads: "The four Ds stay the same. When you do them moves."](img/four-ds-when.png) *Figure 0.4.2. The four Ds in a chat and when the AI works alone.* ## The framework's names for the parts The framework gives each D three named parts.[^anthropic-ai-fluency] Look here when you meet one of these names, in a quiz or a later chapter. | The D | The framework's name | What it means | | --- | --- | --- | | Delegation | **Problem awareness** | Understanding your goal and the kind of work it is, before you bring in AI | | Delegation | **Platform awareness** | Knowing what different AI systems can and cannot do | | Delegation | **Task delegation** | Deciding, task by task, what the AI does and what people do, so that each does what it does best | | Description | **Product description** | What you want back, including its format, its audience and its style | | Description | **Process description** | How the AI should approach the request, such as steps to follow | | Description | **Performance description** | How the AI should behave "during your collaboration",[^anthropic-ai-fluency] such as brief or detailed, challenging or supportive | | Discernment | **Product discernment** | Judging the quality of the output, such as whether it is accurate, suitable, relevant, and whether it makes sense as a whole | | Discernment | **Process discernment** | Judging "how the AI arrived at its output",[^anthropic-ai-fluency] by looking for errors of logic, things it missed and steps that do not belong | | Discernment | **Performance discernment** | Judging how the AI behaved while it worked with you, and whether its way of communicating met your needs | | Diligence | **Creation diligence** | Choosing with care which AI systems you use and how you work with them, such as what you put in | | Diligence | **Transparency diligence** | Being honest about AI's part in your work "with everyone who needs to know"[^anthropic-ai-fluency] | | Diligence | **Deployment diligence** | Taking responsibility for the outputs you use or share, by checking them and standing behind them | [^anthropic-ai-fluency]: AI Fluency: Key Terminology Cheat Sheet, Rick Dakan, Joseph Feller and Anthropic, CC BY-NC-SA 4.0.