How to read this page
This is the plain-English companion to the full AI Fluency guide. Same framework, no jargon.
Every section below has the same four parts, so you can skim whichever one you need:
| Part | What it gives you |
|---|---|
| In one line | The actual definition, no analogy |
| Like you are five | A story you will still remember next week |
| How it goes wrong | The failure mode, which is what makes it stick |
| For a tester | The same idea in your day job |
The definitions are Anthropic's. The analogies are mine, written for this page as teaching aids. Do not quote the sandwich or the babysitter back at an exam; quote the definitions.
The whole thing in one breath
AI Fluency means working with AI in a way that is effective, efficient, ethical and safe. Those four words are the goal. Getting there takes four skills, the four Ds, and the course is careful to draw them as interconnected rather than sequential.
A memory hook for the names: ask, say, check, sign. Delegation is the asking, Description is the saying, Discernment is the checking, Diligence is the signing. Use it to remember the four, but do not mistake it for a running order. Poor Description sends you back to Delegation, and Discernment often tells you the description was the problem. Every pair in that diagram has an arrow going both ways.
D1. Delegation
In one line. Deciding what work to do with AI and what to do yourself.
Like you are five. You have homework and a very fast friend. Some homework you have to do yourself, because the point is that you learn to add up. Some your friend can just do, like copying the list out neatly. And some you do together. Delegation is putting the homework into one of those three piles before you hand anything over.
How it goes wrong. People skip straight to typing. The question "should this be an AI job at all?" never gets asked, so AI ends up doing the one task you actually needed to understand yourself.
For a tester. Should AI write this test, or is reading the requirement carefully the actual job? Generating two hundred boilerplate assertions is pile three. Deciding what "logged in successfully" even means is pile one.
Delegation splits into three. The framework breaks each D into three parts, and for Delegation they are the three things you have to know before you can choose a pile:
| Part | Plain English | The question |
|---|---|---|
| Problem awareness | Know what you are actually trying to do | "What does done look like, in my own words?" |
| Platform awareness | Know what this tool is and is not good at | "Is this even the kind of thing it can do well?" |
| Task delegation | Split the work to suit both of you | "Which bits are mine and which are its?" |
Most bad delegation is a missing first row. You cannot decide whether to hand something over while you are still vague about what it is.
D2. Description
In one line. Communicating effectively with AI systems.
Like you are five. You ask someone to make you a sandwich. "Make me a sandwich" gets you a sandwich, maybe with butter you hate. "Cheese sandwich, no butter, cut into triangles, and I cannot eat nuts" gets you the sandwich you wanted. The other person is not a mind reader and cannot see inside your fridge.
How it goes wrong. You picture the result clearly in your head, describe about a third of it, and then feel let down by the answer. The gap was never in the model. It was in the third you left out.
For a tester. "Write tests for login" is the first sandwich. The second one names the framework, the fixtures that already exist, the acceptance criteria, whether you want the negative cases, and whether assertions belong in the page object.
Description splits into three, and they are genuinely different things to say:
| Part | Plain English | Sandwich version |
|---|---|---|
| Product | What you want back | "cheese, cut in triangles" |
| Process | How you want it worked out | "check the fridge before you start" |
| Performance | How you want it to behave with you | "tell me if we are out of bread, do not guess" |
Almost everyone writes only the first row. Performance is the one people never think to ask for, and it is the one that changes a conversation most: say "challenge my assumptions" or "be brief" and you get a different collaborator, not just a different answer.
D3. Discernment
In one line. Evaluating AI outputs critically.
Like you are five. Your fast friend hands the homework back and the handwriting is beautiful. Neat, straight, lovely. That tells you nothing about whether the sums are right. You still have to check the sums.
How it goes wrong. Fluent writing reads like correct writing. AI output is always well presented, so if you judge by how confident it sounds you will approve wrong answers at a very high speed.
For a tester. The generated test passes. Does it pass because the feature works, or because it asserts something that could never fail? A green test that cannot go red is worse than no test, because now nobody looks at that area again.
Discernment splits into three, and it mirrors Description exactly, which is the neatest thing in the whole framework:
| Part | Plain English | What you are looking for |
|---|---|---|
| Product | Is the answer good? | Wrong facts, missing cases, does not fit the question |
| Process | Was the working sound? | It got there by bad reasoning and landed on the right answer anyway |
| Performance | Is it behaving usefully? | Agreeing with everything, padding, drifting off what you asked |
Product, process, performance in both. Whatever you failed to describe is usually what you now have to discern, which is why the two sit opposite each other on the diagram with an arrow going both ways.
D4. Diligence
In one line. Ensuring responsible AI collaboration.
Like you are five. If you hand in your friend's homework and it is wrong, the teacher does not tell your friend off. Your name is on the page. Diligence is checking it before you hand it in, and being honest when someone asks who helped.
How it goes wrong. Quietly. Nothing breaks at the moment you skip it. It breaks later, in front of other people, and by then the trail back to the decision has gone cold.
For a tester. You approved the pull request. When that suite goes flaky at 2am, "the AI wrote it" is not an answer anyone accepts, and it should not be.
Diligence splits into three, spread across the whole life of the work:
| Part | Plain English | When it happens |
|---|---|---|
| Creation | Choose the tool and the way you use it thoughtfully | Before |
| Transparency | Be honest about AI's part with anyone who needs to know | During |
| Deployment | Check it, then vouch for it | After |
The middle row is the one people quietly skip, and it is the cheapest of the three. A line in the pull request costs nothing and is the whole of transparency.
The three modes, side by side
Now the other half. The four Ds are the skills. The three modes are how much rope you hand over.
Automation, in plain English
In one line. AI executes specific tasks based on your instructions.
Like you are five. A vending machine. You put in a coin, press B4, and out drops exactly that packet of crisps. One button, one thing, and you are standing right there watching it fall.
How it goes wrong. Rarely, and cheaply. That is the appeal. The real risk is doing it fifty times in a row and skipping the check somewhere around number twelve, because eleven were fine.
For a tester. "Convert this one Selenium test to Playwright." You read the diff, you run it, you move on.
Augmentation, in plain English
In one line. You and AI collaborate as creative thinking and task execution partners.
Like you are five. Cooking with a friend. Neither of you makes the whole meal. You taste it, they stir, you say it needs more salt, they adjust. You are both in the kitchen the entire time, and the meal is better than either of you would have made alone.
How it goes wrong. You stop tasting. The conversation keeps going, the answers keep arriving, and at some point you are skimming and nodding rather than thinking. It looks identical to working right up until the moment it does not.
For a tester. Designing the framework together: arguing about whether assertions belong in the page object, sketching the fixture, having your reasoning pushed back on. This is where most real work lives, and it is the mode people skip.
Agency, in plain English
In one line. You guide AI to work independently on your behalf, shaping its knowledge and behaviour rather than specific actions.
Like you are five. You are going out for the evening and a babysitter is coming. You cannot possibly list every single thing that might happen while you are gone. So instead you teach them the rules: bedtime is eight, no sweets after dinner, ring me if there is blood. Then you leave.
Notice what you did there. You never told them what to do. You told them how to decide.
How it goes wrong. Every gap in the briefing gets filled by someone else's judgement, over and over, while you are not in the room. And you find out afterwards, from what is left behind.
For a tester. An agent that triages your nightly CI failures: reads the log, decides flaky or real, files or ignores. You never see most of those decisions. You see the rules you wrote and whatever it left behind.
The order that keeps people out of trouble: automation, then augmentation, then agency. The common mistake is jumping the middle one, because agency sounds like leverage. But augmentation is where you learn what good actually looks like for a task. Write the rules after you know that, not before.
One task, all four Ds
A concrete run through, so the loop is not abstract.
| Step | The question | On a real QA task |
|---|---|---|
| Delegation | Should AI do this? | Two hundred API tests from an OpenAPI spec. Boilerplate, checkable, so yes. Deciding which endpoints matter is mine. |
| Description | Did I say it properly? | Name the framework, the auth fixture, the naming convention, and say I want the 4xx cases too, not just happy paths. |
| Discernment | Is it any good? | Run them. Then break the endpoint on purpose and confirm the tests actually go red. Green means nothing until you have seen red. |
| Diligence | Will I sign it? | Say in the PR that the suite was generated and how it was checked. Then own it when it flakes. |
Try it yourself on the last thing you asked an AI to do. Which of the four did you skip? Almost everyone skips the same one twice: Discernment, because the answer looked finished. Whichever one you skipped is where your next hour of study goes, since your fluency is set by your weakest D and not your average.
The AI words you will hear
The course ships a terminology sheet. These are the same terms in my own words, with the bit that matters to a tester added.
| Term | In plain English |
|---|---|
| Generative AI | AI that makes new things (text, code, images) instead of just sorting or scoring things that already exist. |
| Large language model (LLM) | A generative model trained on an enormous amount of text so it can read and write language. Claude is one. |
| Parameters | The billions of internal numbers a model adjusts during training. They are the model. Nobody hand-writes them. |
| Neural network | Layers of connected nodes that learn patterns from examples. Loosely inspired by brains, not a copy of one. |
| Transformer | The 2017 design that made modern LLMs possible, by letting the model weigh how every word relates to every other word rather than crawling left to right. |
| Scaling laws | The observed pattern that bigger models with more data and compute keep getting better, and that new abilities show up at certain sizes without anyone programming them. An observation, not a law of nature. |
| Pre-training | The first, huge phase where the model learns language and world knowledge from text. |
| Fine-tuning | The follow-up phase that teaches it to follow instructions, be useful, and refuse harmful things. |
| Context window | How much it can hold in mind at once: your conversation plus anything you pasted. There is a ceiling, and it differs by model. |
| Hallucination | Being confidently wrong. The output is fluent and plausible and simply not true. This is the failure mode Discernment exists for. |
| Knowledge cutoff | The date after which the model was not trained on anything. It does not know what happened later unless you tell it or it can look. |
| Reasoning models | Models built to work through a problem in steps before answering, which helps on things that need actual logic. |
| Temperature | A dial for randomness. Higher is more varied and more surprising, lower is more predictable and repeatable. |
| Retrieval augmented generation (RAG) | Giving the model real documents to read at question time, so it answers from sources instead of memory. A direct countermeasure to hallucination. |
| Bias | Systematic skew in outputs that favours or disadvantages some group, usually inherited from the training data. |
Two of these matter more than the rest for testing work. Hallucination is why a confident answer is not evidence, and context window is why a model that was sharp ten messages ago starts dropping details: you have not made it worse, you have run out of room.
The prompting words you will hear
| Term | In plain English |
|---|---|
| Prompt | Everything you send: the instruction plus anything you pasted along with it. |
| Prompt engineering | Getting good at writing those. Mostly clear communication, with a few AI-specific tricks on top. This is Description with a job title. |
| Chain of thought | Asking it to work through the steps rather than jump to an answer. Helps on anything with actual reasoning in it. |
| Few-shot (n-shot) | Showing two or three worked examples of what you want instead of describing it. The "n" is how many examples. Often faster than explaining. |
| Role or persona | Telling it who to be: "answer as a senior SDET reviewing this PR". Changes vocabulary, depth and what it thinks is worth mentioning. |
| Output constraints | Saying the shape up front: how long, what format, which sections. This is the single cheapest fix for disappointing answers. |
| Think-first | Asking for the reasoning before the conclusion, so you can check the working rather than only the verdict. |
Notice that every row in that table is a Description technique. Prompt engineering is not a separate skill sitting next to the 4Ds, it is one of the four with a more marketable name. That is worth knowing before you pay for a course in it.
Where this came from, and the licence. The terminology list is Anthropic's, from the AI Fluency cheat sheet by Rick Dakan, Joseph Feller and Anthropic, released under CC BY-NC-SA 4.0. The definitions above are rewritten rather than reproduced, because that licence is non-commercial and this site sells courses. If you want the original wording, get it from the course itself at Anthropic Academy, which is free.
Where to go next
- The full guide, with Anthropic's own wording and the certificate question answered: AI Fluency: the 4D Framework.
- The other free Anthropic course, module by module: Claude 101 study guide.
- Everything else in one place: the AI for QA hub.