Prompt patterns for real tasks

Most writing requests you will make of an AI tool boil down to four repeatable moves: draft, rewrite, summarize, and change tone or length. Once you recognize which move you need, you stop starting from a blank prompt and start from a pattern that already works.

The four moves

• Draft: generate new text from a description, when you are starting from nothing.
• Rewrite: improve existing text while keeping its meaning the same.
• Summarize: compress a longer piece down to its essential points.
• Change tone or length: keep the same content but reshape it for a different audience or space.

Paste the text, don't describe it

For rewrite, summarize, and tone or length changes, paste the actual text into the prompt instead of describing it. "Make my email more polite" gives the model nothing to work with. Pasting the email and adding "make this more polite, keep the meaning the same" gives it the exact material to edit.

Code tasks follow the same idea as writing tasks: a handful of moves cover most requests. The four common ones are explain, generate, debug, and refactor. What separates a fast, correct answer from a frustrating back-and-forth is not the wording of the request. It is the context you give it.

The four moves

• Explain: paste a function or block and ask what it does and why, or what a specific line is for.
• Generate: describe the inputs, outputs, and edge cases you need, not just the goal.
• Debug: paste the code, the exact error message or wrong output, and what you expected instead.
• Refactor: paste the code and say what must stay the same (the behavior) and what should change (readability, speed, structure).

The key move: precise context

The single biggest lever in a code prompt is precision. "My function isn't working" tells the model almost nothing; it has to guess at the language, the bug, and what "not working" even means. Pasting the real code and the real error removes the guessing entirely.

This Python function should return the average of a list, but it
returns the wrong number.

def average(nums):
    total = 0
    for n in nums:
        total = n
    return total / len(nums)

average([2, 4, 6]) returns 2.0. I expected 4.0.

What is the bug, and what is the fix?

Beyond writing and code, AI is useful for turning messy text into something structured you can actually use. Three moves cover most of this work: extraction, classification, and structured output.

The three moves

• Extraction: pull specific fields out of unstructured text, like names, dates, or amounts, into a list.
• Classification: sort items into categories you define, such as urgent versus not urgent.
• Structured output: ask for the answer in an exact shape, like bullets, a table, or JSON, so it slots into whatever comes next.

Naming the shape you want

The output format matters as much as the content. If you just ask for "the info," you get a paragraph you have to pick apart by hand. If you ask for a table with named columns, or JSON with named keys, you get something you can paste straight into a spreadsheet or another program.

Extract the following fields from the text below and return them
as JSON with exactly these keys: name, date, amount.

Text: "Invoice #4421, billed to Sarah Chen on March 3, 2026,
for $1,250."

Return only the JSON object, no other text.

The first result you get from a prompt is rarely the final one. Good prompting is not a spell you get right on the first try. It is a short loop you run a few times.

The refine loop

Why one thing at a time

If you change the tone, the length, and the format all in the same edit and the new result is better, you do not actually know which change fixed it. Change one variable, compare the two outputs side by side, and you learn what that one thing controls. Do this a few times and you build real intuition for how a model responds to your prompts, instead of guessing at a new prompt from scratch every time.

What counts as "one thing"

• Adding one example of the output you want.
• Making the audience or purpose more specific.
• Asking for a different format, bullets instead of prose, or a table.
• Removing an instruction the model keeps ignoring.
• Shortening or lengthening the ask.