The Prompt Cheat Sheet

Everything you need to write a good prompt, on one page. Keep it next to your keyboard, or download it and paste the template into your notes.

The master template

Fill in the brackets and delete the lines you don't need.

You are [ROLE].
Task: [WHAT TO PRODUCE].
Context: it is for [AUDIENCE], who [WHAT THEY KNOW / NEED]. It matters because [WHY].
Use only these facts: [FACTS, IF ANY].
Constraints: [LENGTH], [TONE], avoid [WHAT TO AVOID].
Format: [STRUCTURE — e.g. 5 bullets, a table with columns X/Y/Z].
Example of what good looks like: [EXAMPLE, OPTIONAL].
If anything is missing or unclear, ask me before answering.

The six parts of a strong prompt

Most weak answers are missing one of these. You rarely need all six; check which one is absent.

Role: Who the AI should act as — "a patient maths tutor", "a B2B copywriter".
Task: The exact thing to produce, in one sentence.
Context: Who it is for, the situation, and why it matters.
Constraints: Length, tone, what to avoid — each with a short reason.
Format: The shape of the answer: bullets, table, sections, word count.
Examples: One to three samples of what good looks like, in the same format.

Frameworks at a glance

Pick one when you are stuck on structure. Each has a full guide on PromptVibe.

ROLE: Role, Objective, Limitations, Expectations — expert personas, professional writing.
TAG: Task, Action, Goal — quick, direct requests.
COAST: Context, Objective, Actions, Scenario, Task — strategy and full-context briefs.
RISE: Role, Input, Steps, Execution — step-by-step workflows and tutorials.
APE: Action, Purpose, Execution — data tasks and process automation.
RACE: Role, Action, Context, Explanation — teaching and explanations.

Fix-it follow-ups

Name the problem and say what must stay the same. Never just "make it better".

Too generic: "The reader is [WHO] in [SITUATION]. Rewrite it for them using: [SPECIFICS]."
Too long: "Cut to [N] words. Keep every name, date and number exactly as written."
Wrong tone: "Same content, but sound like [DESCRIPTION]. Do not add new claims."
Missed a rule: "You missed: [RULE]. Fix only that, then show where each rule is met."
Made-up facts: "Remove anything not in the facts below; mark gaps as [CHECK]."
Wrong format: "Same information as [FORMAT]. Do not add or remove anything."

Model-specific tips (from the providers’ own guides)

Claude: Wrap instructions, documents and examples in XML tags; put long documents first and the question last; explain the reason behind each rule.
Claude (API): Prefilling the reply is not supported on the newest models — ask for the format directly or use structured outputs.
Gemini: Always include a few examples, formatted identically; use one delimiter style; put long context first and the question at the very end.
Gemini: Current models reason internally — you usually do not need to ask for step-by-step reasoning in the answer.
Any language: Name the output language and its variety (e.g. Mexican Spanish, Canadian French, Traditional Chinese), write in native script, and ask for localisation, not translation.
Image models: Subject → setting → style → composition → lighting → aspect ratio. Put the most important element first.

Before you send: a 6-point check

1: Could a stranger do this task from my prompt alone?
2: Did I say who it is for?
3: Is there a length or format target?
4: Did I give the facts, instead of hoping the AI knows them?
5: Did I say what to do if information is missing?
6: Will I check names, numbers and claims before using the answer?

Skip the template entirely

Describe your goal in one line and the free generator fills in all six parts for you, tuned for ChatGPT, Claude, Gemini or an image model.

Open the free generator →

Model tips summarise Anthropic's and Google's published prompting guides as of October 2026. Read the full guides: Claude · Gemini · frameworks.