What Anthropic, OpenAI and Google Say About Prompting

Study guide

What the AI Makers Say About Prompting

Anthropic, OpenAI and Google each publish their own prompting guidance, and most people never read any of it. This course puts all three side by side and shows you where they agree. The short answer is that all three want clear prompts with real context, a few examples and a defined outcome, and none of them asks you to memorise a rigid formula. You will learn Google's four prompt areas, Persona, Task, Context and Format, Anthropic's golden rule and its habit of wrapping prompt parts in XML tags, and OpenAI's outcome first approach with success criteria and constraints.

The course is built for busy professionals who use ChatGPT, Claude or Gemini at work and want to prompt better without learning three separate rulebooks. Every lesson explains the why, the how, a worked example in a real job and a common mistake. You will see why modern reasoning models need fewer rules and more judgement, why over-prompting with demands and capital letters makes models over-eager, and how to match your prompt to the assistant you are actually using. By the end you will have one repeatable habit that works across all three tools.

What is a prompt framework?

A prompt is the request you type into an AI tool such as ChatGPT, Claude or Gemini. A prompt framework is a reusable structure for that request: a short checklist of the parts a good prompt contains, usually named with an acronym so it is easy to remember. Why it helps: the AI fills in whatever you leave out with generic guesses. A framework makes sure you include the parts that change the answer, such as who it is for, what you want to achieve and the format you need.

Frameworks make prompts repeatable: you fill the same parts each time, save the prompt as a template and share it with colleagues. They are not magic words: the value comes from the detail you put in each part. They help most when a request is complex or open to interpretation; a quick, simple question can stay short. There are well over sixty named frameworks, and most of them remix the same ingredients: role, task, context, audience, format, constraints and examples.

In this course you learn Official guides, which is the official prompting advice from the companies that make the AI; the frameworks build on the same principles. It is best for: anyone who uses Claude, ChatGPT or Gemini and wants the makers' own rules.

Three companies, one message

Anthropic, OpenAI and Google all publish prompting guides, and none of them prescribes a single rigid formula. The shared principles are clarity, context, examples and defining the outcome, then letting the model reason. A strong prompt names the role, the task, the context and the format. That structure comes from Google's guide, and it works in ChatGPT, Claude and Gemini alike. You do not need three different rulebooks. The overlap between the guides is the practical skill, and it transfers to any tool you pick up.

The newest guides all say the same thing about length. Modern models need fewer rules and more judgement, so you can describe the situation and the goal rather than writing a long list of instructions. Anthropic cut over 80 percent of the system prompt for Claude Code when working with Claude 5 models, with no measured loss in performance. That is a useful signal. Long prompts are often doing less work than people think, and the effort spent maintaining them could go into context instead.

This matters in ordinary jobs. If you write a nine step procedure for turning meeting notes into a client update, you spend your time on steps the model already handles. If you write a short brief with the audience, the goal and the tone, you spend your time on the part only you know. The rest of this course takes each company's advice in turn, then shows how to combine it into one habit.

Google: Persona, Task, Context, Format

Google's Gemini for Workspace guide names four areas of a prompt: Persona, Task, Context and Format. Treat them as a checklist, not a required form. Google notes you rarely need all four at once, and adding a slot does not improve a prompt by itself. A persona only helps when tone or role matters. Context only helps when the AI cannot guess the facts. Add a slot when it changes the answer, and leave it out when it does not.

The verb is the part Google calls most important. It tells the AI what action to take, such as summarise, draft, compare or list. A prompt without a verb leaves the intention to guesswork, and you get an answer to a question you did not ask. Google also found that the average prompt is about nine words, while the most effective ones are around twenty-one. The extra length carries the persona, context and format that make an answer usable rather than generic.

In practice the four areas map to real prompt slots. Persona sets the role, Task carries the verb, Context supplies the facts, and Format shapes the answer. Suppose you need to turn a messy project update into something a client can read. Persona: you are an account manager. Task: rewrite this update as a short email. Context: the client is non-technical and wants progress and risks. Format: three short paragraphs, no bullet lists. That is twenty-one words of direction and it beats nine words of hope.

Anthropic: the golden rule and structure

Anthropic's golden rule is a test for prompt clarity. Show your prompt to a colleague with minimal context. If they would be confused, Claude will be too. The fix is to write out the context you were holding in your head, the background you assumed everyone shared. Anthropic also asks you to explain the why behind each instruction. A reason such as the reader is a busy client lets Claude make sensible choices on details you did not spell out.

Anthropic recommends three to five examples of the output you want. Fewer leaves Claude guessing and more adds weight without much gain. It also uses XML tags, which are simple labels in angle brackets that wrap each part of a prompt, for example documents, task and rules. Tags keep your instructions separate from your source material, so Claude does not confuse the two. Giving Claude a role sets the voice and the standard for the answer. A role such as legal assistant tells Claude which knowledge and tone to bring.

One structural habit is easy to miss. Put long documents at the top and your question at the end. Anthropic reports up to thirty percent better results on long inputs with this order. If you paste a contract and then ask your question, the question lands right after the material it refers to, which is where the model is paying attention. Reverse the order and the model reads your question before it has the facts.

OpenAI: outcome first

Outcome first prompting means you describe what you want to exist at the end, not the steps to get there. The model chooses the path, which often produces a better result than a rigid process you invented. Success criteria are how you will judge the answer. If you do not state them, the model guesses. Stating them up front saves a round of corrections, because the model knows what good looks like before it starts writing.

Contradictory instructions waste effort because precise models follow them literally. If two instructions clash, the model tries to satisfy both and the answer suffers. Read your prompt back and remove clashes. OpenAI's GPT-5 era guides recommend this structure: define the target outcome, success criteria, constraints and context, then let the model choose the path. It works best when the instructions do not fight each other.

Two settings carry a lot of weight. Reasoning effort controls how much thinking the model does before answering, so raise it for hard problems and lower it for simple ones. Verbosity controls how long the answer is, so match it to what you actually need. You can also ask the model itself to improve your prompt. Paste your prompt and ask what is unclear, what is missing, and what a stronger version would look like. Then use its suggestions.

What changed with reasoning models

Reasoning models plan and check their own work, so instructions like think step by step add little. The model already does that internally, and repeating it can distract from the real task. Over-prompting means pushing too hard with demands, threats or capital letters. Anthropic advises dialling back aggressive instructions because they make models over-eager, guessing or rushing to please. A calm brief gets better work than a shouted one.

Context is now the main lever. A reasoning model can fill gaps, but only with guesses, so your real situation, limits and goal set the quality of its thinking. A clear outcome tells the model what good looks like without scripting every step. State the goal, the constraints and the format, then let the model reason about how to get there. This is the calm brief pattern: role, task, context, constraints and format give the model enough to work with.

One addition makes the calm brief stronger. Invite the model to ask questions. A line such as ask me if anything is unclear catches gaps early, before the model invents an answer. In a real job this is the difference between a report built on the right numbers and a report built on a plausible guess. The invitation costs one sentence and saves a rewrite.

Applying the official advice

The colleague test asks whether a new colleague could complete the task from your prompt alone. If they would need to ask a question, that missing detail belongs in the prompt. It is the fastest way to catch vague instructions before you send them. A strong prompt holds four parts: outcome, context, constraints and one example. The outcome is what good looks like, the context explains who it is for and why, the constraints set length and tone, and the example shows the format you want.

Leave the method to the model unless the method matters. If the steps must happen in a set order, say so clearly. If not, let the model choose its own route, because modern models often plan better than a fixed list you paste in. Matching the prompt to the tool matters too. A long method on a reasoning model can get in the way, and a deep task on a quick model will miss detail. Match the prompt to the assistant and the depth the task needs.

Each assistant has features that carry part of the work. Projects keep related chats and files together, custom instructions carry your role and tone into every chat, effort settings control how deep the answer goes, and connected tools let the model read your files or workspace. The course comes together in one habit. Write for a colleague, lead with the outcome, add context and constraints, give one example, leave the method open unless it matters, then use the assistant's own features. That habit works across ChatGPT, Claude and Gemini.

Frequently asked questions