Prompt Academy · 7 lessons
How to write a better AI prompt
Seven parts, one lesson each, every one ending in a live sandbox that scores what you wrote. Work through them in order and the seven bars stop being a mystery , or read the whole rubric in one go further down.
The curriculum
- 01Role prompting: how to tell an AI who it isThe single fastest way to raise output quality: tell the model who it is before what to do.Beginner · 3 min
- 02How to define the task in an AI promptThe heaviest-weighted thing the scorer looks for, and the one most prompts fumble: an explicit action.Beginner · 3 min
- 03How much context to put in an AI promptStop the back-and-forth. Put every fact the model needs into the first message.Beginner · 4 min
- 04How to control an AI's output formatTell the model the exact shape of the answer so you never have to re-ask for formatting.Intermediate · 3 min
- 05Prompt constraints: rules an AI will followDo/don't rules keep the model on-scope and prevent hallucinated extras.Intermediate · 3 min
- 06Few-shot prompting with a single exampleA single input→output example is the cheapest reliable accuracy boost.Advanced · 4 min
- 07Why vague prompts fail, and how to fix themEvery vague word is a decision you handed to the model. Take them back.Intermediate · 3 min
The seven parts, in order
A prompt that works is not a longer prompt. It is one that has answered seven questions before the model has to guess at them. Below is the whole rubric in the order the parts go in: the same seven the analyser scores, and the same seven the lessons above teach one at a time.
01Give it a role
A role primes tone, vocabulary and standards before the model reads a single instruction. “You are a senior security engineer” pulls in caution, threat-modelling and precision that “help me with this code” never will. Be specific about seniority and domain: “senior B2B copywriter” beats “writer”. Every production agent worth copying opens by establishing exactly who it is.
Role prompting: how to tell an AI who it is \u00b7 3 min \u219202Name one action
This is the heaviest-weighted part of the score and the one most prompts fumble. Lead with one imperative verb and its object: refactor this function, draft this email, classify these tickets. One verb, not three. A prompt that asks to research, summarise and then design is three prompts wearing a trenchcoat. The test: read only your first eight words. If someone could do the wrong job from them, the verb is buried.
How to define the task in an AI prompt \u00b7 3 min \u219203Front-load the facts
Missing context is the main reason a prompt turns into a conversation: the model asks, you answer, repeat, and every round costs time. Put what the model needs in the first message: who the audience is, what data it has, what a good result looks like. The rule is not “more context”. It is that any fact which would change the answer belongs in the prompt.
How much context to put in an AI prompt \u00b7 4 min \u219204State the shape of the answer
Without a stated format you get prose where you wanted JSON, or three paragraphs where you wanted a table. Say the shape exactly: “Return JSON: { subject, body, cta }”, or “a markdown table with columns name, owner, due date, under 200 words”. This is also what makes a prompt reusable, because a specified output can be parsed, diffed and passed to something else.
How to control an AI's output format \u00b7 3 min \u219205Say what not to do
Constraints stop the model wandering into work you did not ask for. Length limits, scope limits, and explicit do-not rules: “Keep it under 150 words. Do not invent APIs. Use only the data provided.” They earn their place on the second run, when the model has room to be inventive about a detail you had already settled.
Prompt constraints: rules an AI will follow \u00b7 3 min \u219206Show one worked example
A single input-to-output example removes more ambiguity about tone, format and edge cases than a paragraph of description does. It is the cheapest accuracy gain available, and one strong example reliably beats three weak ones. Show a representative input and the exact output you would have accepted for it.
Few-shot prompting with a single example \u00b7 4 min \u219207Delete the hedges
Every vague word is a decision handed back to the model. Some, various, appropriate, properly, etc. read like instructions but carry no information, so the model guesses, and its guess is an average of everything it has read, which is the generic output you were trying to avoid. Trade each hedge for a number, a name or a rule: “keep it short” becomes “under 120 words”.
Why vague prompts fail, and how to fix them \u00b7 3 min \u2192The seven are not equally weighted. Task carries the most, because a model that cannot tell what job it is doing gets everything after it wrong too. Role and context come next. Examples matter least in isolation and most when the other six are already in place, which is why the curriculum runs in this order rather than by difficulty.
Reading about it is the slow half. Paste a prompt you actually sent this week into the desk and it will mark which of the seven you left out. Or take today\u2019s drill . One brief, one prompt, scored against everyone else who tried it.
Writing something specific? 8 guides apply the whole rubric to one job: cover letters, SQL, code review, debugging, and the templates hand you the same seven parts already filled in.