Everyday Prompting Cookbook

01. What It Is

A prompt cookbook is a set of reusable patterns you fill in, not scripts you memorize.
This page is the applied companion to two foundation files, and it sits under how-to-use-an-llm, the home base for everyday use. prompting-basics explains what a prompt is and what zero-shot and few-shot mean, and prompt-engineering-patterns covers production techniques like chain-of-thought. This file does neither. It hands you a simple shape, a few reusable ingredients, and task templates you can copy, adapt, and use. When a recipe brushes against why a technique works, it links the theory rather than repeating it.

02. Why It Matters

The fix for most weak prompts is one mental shift. Treat the tool like a new teammate you are training, not a search box. Google's guide makes the point with a bad example. You wouldn't tell a new hire "Sales report, please!" and walk away, you would give context and suggest a format. Anthropic reaches the same image on its own, a brilliant but new employee who lacks context on your norms, so the more precisely you explain what you want, the better the result. OpenAI puts it as writing to the model the way you would write to a person.

The numbers say the same thing. Google's research found the most useful prompts average about 21 words, while people's first attempts usually run under 9. The lesson is not "write more," it is "add the missing context and constraints."

03. How It Works

The shape below is the spine. Persona, task, context, format, and always a verb. You don't need all four parts every time, but adding a few will sharpen almost any request.

The core ingredients of a good prompt

Start with the four-part shape Google teaches for normal users, known as PTCF.

• Persona is the perspective the AI should adopt, such as "You are a friendly HR manager." It shapes voice, not facts.
• Task is the specific action, and it must contain a verb. Google calls the verb "the most important component of a prompt." Reach for summarize, draft, rewrite, compare, or list.
• Context is the background the AI needs, including who the answer is for, the goal behind it, and the actual text to work on. This is the most common missing ingredient.
• Format is how the answer should look, such as a bulleted list, a table, a three-sentence paragraph, or an email draft.

Here is the shape with each part labeled inline. "I'm a project manager (persona) and need a project tracker (task) for a website redesign (context), as a simple table with columns for date, status, and task (format)."

Now add ingredients on top of the shape.

• Be specific about length and format. A number beats "short." OpenAI's own fix replaces "a few sentences only" with "a 3 to 5 sentence paragraph." Words like "a few," "detailed," and "some" are guesses you hand the model.
• Assign a role to set tone and expertise. Anthropic notes that even a single sentence of role makes a difference, and Google suggests roles like "You are the head of a creative department." A role narrows vocabulary and assumed knowledge, it does not add facts.
  See system-prompts.
• Show an example of what good looks like. OpenAI sums it up as "show, and tell." Start with no example, then add one or two only if the format is not coming out right. Anthropic calls examples one of the most reliable ways to steer output and suggests 3 to 5.
  See prompting-basics.
• Say what to do, not only what not to do. Replace "do not ask for personal details" with "refer the user to the help article instead." A prohibition on its own leaves the model no target to aim at.
• For hard tasks, ask it to work step by step. Anthropic's guidance is that giving the model space to think reduces errors on harder reasoning like math and logic. Use it for analysis, not simple lookups.
  The depth lives in prompt-engineering-patterns.

One check ties these together. Anthropic offers a golden rule. Show your prompt to a colleague who has minimal context and ask them to follow it. If they would be confused, the model will be too.

Iterating and refining

The follow-up sentence is the whole skill. The chat keeps your earlier context, so you refine the answer instead of rewriting the prompt. All three providers describe prompting as a conversation. OpenAI says to refine your requests based on the first answers, and Google says to "say it another way" and to keep asking follow-up questions.

These one-line follow-ups cover most needs.

• "Make it shorter." / "Make it longer."
• "Make it warmer." / "Make it more formal."
• "Use bullet points instead." / "Now as a table."
• "Rewrite it for a non-technical reader."
• "What did you assume that I should correct?"

Two more habits help. Keep one job per prompt, since Google advises breaking several related tasks into separate prompts, and cramming several jobs into one request degrades all of them. And you can let the model interview you first. Hand it the project details, then ask "What questions do you have for me that would help you provide the best output?"

Asking for sources and saying "I don't know"

A confident answer can still be wrong, so two instructions help you catch it. Give the model permission to be unsure, and ask it to back its claims. Anthropic's guidance for normal use is to allow the model to say "I don't know," which "can drastically reduce false information," and to have it cite quotes and sources for each claim. When you paste in a document, tell it to use only that document and not its general knowledge. Then do your part and open the sources it cites.
See fact-checking-ai-answers.

04. Recipes

Each recipe is a template. Copy it, replace the brackets, and paste your own text where marked. Start simple, then refine with a follow-up.

Summarize a thread or document.

Summarize the text below in [5] bullet points, then list any action items and who owns them.
Text: """[paste the email thread or document]"""

Rewrite or change the tone.

Rewrite the text below in a [friendly, professional] tone, in plain English, no longer than [120] words.
Text: """[paste your draft]"""

Draft an email.

Write an email to [recipient] to [your goal, for example negotiate a bulk discount]. Use a [collaborative] tone and keep it under [150] words.

Brainstorm with pushback.

Give me [10] ideas for [your goal]. Add one line of pros and cons for each, and include a few ideas I may not have considered.

Explain it simply.

Explain [topic] to a [smart 12-year-old / busy executive] in [3] sentences, with no jargon.

Extract and structure information.

Pull the [name, date, amount, due date] out of the text below into a table.
Text: """[paste the messy text]"""

Compare options.

Create a table comparing [option A] and [option B], with columns for [price, timeline, and fit].

Plan or make a checklist.

Turn this goal into a step-by-step plan with a checklist: [your goal and any constraints]. Then ask me what questions would help you improve it.

Two fixes from the providers' own guides show the pattern. "Write a poem about OpenAI" becomes "Write a short inspiring poem about OpenAI, focusing on the recent DALL-E product launch, in the style of [a famous poet]," which supplies the angle and style the first version left blank. "The description should be fairly short, a few sentences only" becomes "Use a 3 to 5 sentence paragraph to describe this product," which swaps a fuzzy size word for a number.

05. Key Terms

06. Common Misconceptions

"There is one perfect magic prompt, and pros know the secret words."
No. The providers' own guidance is about clear, specific requests with enough context, not incantations. Google states it outright, you don't have to be a prompt engineer. The reliable lever is context and constraints, plus iteration, not a hidden phrase.

"Longer prompts are better."
Not length, fit. Good prompts average about 21 words with relevant context, while first attempts run under 9. You are usually adding the missing context and format, not padding, and a short prompt with a clear task beats a long ramble.

"Listing what I don't want is enough."
A prompt made only of prohibitions gives the model no target. OpenAI's rule is to say what to do instead. "Write 3 short, friendly sentences in plain English" works better than three "don'ts."

"If the first answer is off, the tool just can't do it."
Iteration is the expected workflow, not a sign of failure. The fix is usually a one-line follow-up like "make it warmer" or "now as a table," and the chat keeps your earlier context so you do not start over.

"A confident, well-formatted answer is a correct answer."
Format is not truth. These models can state false things fluently. This is why the providers themselves suggest letting the AI say "I don't know" and asking it to cite sources, and why you review before you use it.
See fact-checking-ai-answers and hallucination-grounding-guardrails.

"Telling it to 'act as an expert' makes it know more."
A role changes tone and assumed audience, not the underlying facts. "You are a tax attorney" makes the answer sound like a tax attorney, it does not make it correct or current. Use roles for voice, and still verify the substance.
See system-prompts.