Prompting Cheat Sheet

One page to keep open beside your AI assistant: the six-part recipe, twenty-one copyable templates for everyday tasks, quick fixes when an answer disappoints, and the things you should never paste.

You do not need secret phrases or a paid prompt pack. Nearly every good prompt is the same few ingredients in a sensible order, and nearly every bad answer has one of a handful of causes. This page is a reference, not a lesson: scan it, copy a template, replace the [SQUARE BRACKETS] with your own details, and go. For the reasoning behind it, follow the links to Prompting Basics, Prompt Craft and Working with Context.

The core recipe

Consensus The official prompting guides from Anthropic, OpenAI and Google agree on the essentials: be clear and specific, give context, show examples, say what format you want, and state the limits. You rarely need all six parts; for anything that matters, use at least the first three.

PartAsk yourselfExample phrase
TaskWhat exactly should it do?“Draft a reply to…”, “Compare…”, “List…”
ContextWhat does it need to know that it cannot guess?“We are a 20-person clinic; the patient asked…”
AudienceWho will read the result?“for parents with no technical background”
FormatWhat shape should the answer take?“a table with columns…”, “3 bullets”, “under 150 words”
ExamplesCan I show what good looks like?“In the style of this sample: …”
ConstraintsWhat must it avoid or keep?“Keep all figures unchanged; do not invent names.”

Three habits make any recipe work better. Put long material at the end inside clear markers such as <text> … </text>. Invite questions: “If anything is unclear, ask me first.” And iterate — your second message (“shorter”, “more formal”, “you missed X”) is often worth more than a perfect first one. Showing examples is called few-shot prompting; standing preferences you want every time belong in custom instructions rather than in each prompt.

Task only. Fine for a quick fact, an idea or a definition, where you will judge the answer yourself.

Quick ask
Give me five name ideas for a [TYPE OF EVENT] for [AUDIENCE].

Everyday writing

1 · Email from notes
Turn my notes below into a clear email to [RECIPIENT ROLE]. Purpose: [WHAT I NEED FROM THEM] by [DATE]. Tone: [polite and direct / warm / formal]. Under [N] words, with a subject line. Do not add facts that are not in my notes.

<notes>
[YOUR NOTES]
</notes>
2 · Polite reply, firm message
Help me reply to the message below. I want to [decline / say not yet / ask for changes] while keeping the relationship good. Be polite but unambiguous; no over-apologising. Give me two versions: one short, one slightly warmer.

<message>
[PASTE THE MESSAGE, WITH NAMES REMOVED]
</message>
3 · Rewrite in simpler words
Rewrite the text below so that [AUDIENCE, e.g. a 12-year-old / someone new to the topic] can understand it. Use short sentences and everyday words, explain any term you keep, and do not change the meaning. Then list anything you had to simplify so much that it is no longer exactly accurate.

<text>
[TEXT]
</text>
4 · Translate BM ↔ English
Translate the text below from [Bahasa Melayu / English] into [English / Bahasa Melayu] for [AUDIENCE, e.g. a government office / customers / friends]. Keep the tone [formal / casual]. Keep names, numbers and dates exactly. After the translation, list any phrases where a literal translation would sound wrong and explain your choice.

<text>
[TEXT]
</text>
5 · Edit, don’t rewrite
Edit my draft below for clarity and grammar only. Keep my voice, my structure and my word choices wherever they work. Show the edited version, then a short list of the changes you made and why.

<draft>
[YOUR DRAFT]
</draft>

More on keeping your own voice: AI for Writing.

Learning

6 · Explain like…
Explain [TOPIC] to [a curious 12-year-old / a new employee / someone who knows [RELATED FIELD]]. Start with an everyday analogy, then give the accurate version in three short paragraphs. End with one common misconception and why it is wrong.
7 · Quiz me
Quiz me on [TOPIC / the notes below] at [beginner / intermediate] level. Ask one question at a time and wait for my answer. After each answer, tell me if I was right, explain briefly, and adjust the difficulty. After 8 questions, summarise what I should review.
8 · Socratic tutor
Act as a patient tutor helping me understand [TOPIC / PROBLEM]. Do not give me the answer. Ask me guiding questions one at a time, give a small hint only when I am stuck, and let me reach the conclusion myself. If I state something wrong, ask a question that helps me notice it.
9 · Study plan
I want to learn [SKILL] in [N] weeks with about [N] hours a week. My current level: [DESCRIBE]. Make a week-by-week plan with one practical exercise per week and a way to check my progress. Prefer free resources, and say “check this exists” next to any resource you name.

Asking to be quizzed uses retrieval practice, one of the best-supported study methods. More in AI for Research and Learning.

Work

10 · Summarise a document
Summarise the document below for [AUDIENCE] who need to [DECIDE / ACT ON] it. Give: (1) a three-sentence summary, (2) the key points as bullets, (3) any deadlines, figures or obligations, quoted exactly, (4) anything unclear or missing. Use only the document; say “not stated” rather than guessing.

<document>
[PASTE APPROVED, NON-CONFIDENTIAL TEXT]
</document>
11 · Meeting notes → actions
From the meeting notes below, produce: (1) decisions made, (2) a table of action items with columns Task, Owner, Due date, (3) open questions. If no owner or date is stated, write “not stated” — do not invent them.

<notes>
[NOTES]
</notes>
12 · Compare options in a table
Compare these options for [GOAL]: [OPTION A], [OPTION B], [OPTION C]. Make a table with columns: Option, Cost, Effort, Main benefit, Main risk, Best for. Use only the information I give below; mark gaps as “unknown”. Then give a two-sentence recommendation and say what would change it.

[DETAILS OF EACH OPTION]
13 · First draft of a policy
Draft a first version of a [TOPIC] policy for [ORGANISATION TYPE, SIZE]. Headings: Purpose, Scope, Rules, Responsibilities, Exceptions, Review date. Plain language, under [N] words. Mark with [CHECK] every point that depends on law or our own rules, so a qualified person can confirm it. Do not cite laws by section number unless you are certain.
14 · Spreadsheet formula
In [Excel / Google Sheets], column [A] has [WHAT], column [B] has [WHAT], data runs from row 2 to row [N]. Write a formula for [RESULT I WANT]. Explain it in plain words and give one example row with the expected result so I can test it.

Long documents and files: see Working with Context. Numbers: AI for Data and Spreadsheets.

Thinking

15 · Pros and cons
I am deciding whether to [DECISION]. My situation: [CONTEXT]. My priorities, in order: [1, 2, 3]. List the strongest pros and cons, weigh them against my priorities, and tell me which questions I should answer before deciding. Do not decide for me.
16 · Devil’s advocate
Here is my position: [POSITION]. Argue against it as strongly and fairly as you can, the way a well-informed critic would. Give the three best counter-arguments, then say which one I most need to answer.
17 · Find the flaws in my plan
Review my plan below as a sceptical, experienced [ROLE, e.g. project manager]. Identify the five biggest risks or weak assumptions, rate each high / medium / low, and suggest one practical fix for each. Be direct; do not soften the criticism.

<plan>
[PLAN]
</plan>
18 · Think it through first
Before answering, list what you know, what you are assuming and what information is missing for this question: [QUESTION]. Then give your answer, and say how confident you are and why.

Assistants tend to agree with you (sycophancy). Templates 16 and 17 deliberately push the other way.

Checking

19 · List the claims to verify
List every factual claim in your previous answer as bullet points. For each, say how confident you are (high / medium / low) and what kind of source would confirm it. Flag any number you did not calculate and any source you cannot be sure exists. Do not add new claims.
20 · Sources and uncertainty
Answer [QUESTION]. For each key point, name the type of source it comes from and, if you can, the specific official source. Separate clearly what is well established, what is debated, and what you are unsure of. If you do not know, say so rather than guessing.
21 · Check against the source
Compare your summary with the original text below. List anything in the summary that is not supported by the text, anything important the summary left out, and any figure that differs. Quote the original for each point.

<original>
[TEXT]
</original>

These prompts help you find where to look; they do not do the looking. Open the sources yourself. The full routine is in Checking AI’s Work, and a hallucination can survive a model’s own review.

When the answer is bad

This guide’s framing Most disappointing answers trace back to a missing part of the recipe. Diagnose, then send a follow-up rather than starting over.

SymptomLikely causeFollow-up to send
Generic, blandNo context or audience“This is for [AUDIENCE] who [SITUATION]. Make it specific to that.”
Too longNo format or length“Cut to [N] words / 3 bullets. Keep only what [AUDIENCE] needs to act.”
Wrong toneNo example“Match the tone of this sample: [SAMPLE].”
Invented factsAsked for facts it does not have“Use only the text I gave you. Write ‘not stated’ where it is missing.”
Misunderstood the taskAmbiguous ask“Before answering again, restate my request in your own words and ask me what is unclear.”
Agrees with everythingLeading question“Give me the strongest case against this.”
Out of dateKnowledge cutoffTurn on web search, or paste the current official source and say “answer from this”.
Long chat going off trackToo much old contextStart a new chat and paste a short summary of what matters.
Diagnose it

You asked “Write a speech for our event” and got something that would suit any event anywhere. Best follow-up?

Generic answers mean missing context and audience. “Make it better” gives the model nothing new to work with.

A summary of your report includes a figure that is not in the report. What do you send?

Constrain it to the source, then check the figures yourself. “Are you sure?” often just produces confident agreement or a new guess.

You ask about a regulation and the answer describes rules from two years ago. Why, and what now?

Models only know what was in their training data unless they search or you supply the source. For rules, the official text is the authority anyway.

Never paste

The privacy rule

Templates above say [PASTE] — but only paste what you would be comfortable seeing on a public notice board. Never paste: IC or passport numbers, bank or card details, passwords, one-time codes or API keys; other people’s personal data (names with health, financial or contact details); confidential work documents, client data or unreleased plans unless your employer has approved that tool for it; medical records. Remove or replace names and identifiers first (“Client A”, “[AMOUNT]”). Check your workplace’s AI rules, and remember Malaysia’s PDPA applies to personal data you handle. More in Staying Safe with AI and AI in the Workplace.

One more safety habit: be wary of pasting text from unknown web pages or emails and asking the assistant to “follow” it. Hidden instructions in such content can steer the model — a risk called prompt injection.

// Pause and reflect

  1. Which three templates match tasks you do every week? Save your own filled-in versions somewhere you can find them.
  2. Think of your last disappointing AI answer. Which part of the recipe was missing?
  3. Is there anything you have pasted into an AI tool that, looking at the “never paste” list, you should not have?

Your notes are saved privately in this browser.

Try this

Pick one real, non-confidential task from your week. Run it once as a one-line request, then again with template 1, 10 or 12 filled in properly. Compare the two answers and note how many follow-ups each needed.

Key terms on this page

Sources & further reading