steve sparks
28 of 53
ContentsPart IV · Chapter 13

Chapter 13Asking well

The parts of a request

In this chapter How a good request is built: the job, who it’s for, background, an example, limits and the shape you want.

Words to know

Prompt
Your message. More broadly, everything sent to the model.
Context window
How much the model can "see" at once. The size of its desk, measured in tokens.

A good request says the job, who it’s for, the background, an example, the limits and the shape you want. Whatever you leave out, the model has to guess.

Chapter 8 showed that your message is only one layer of what the model reads. It is the layer you control. The companies that make these models give the same advice in their own guides: say clearly and specifically what you want, explain the context or the reason behind it, show an example, and say what form the answer should take.12

What would it have to guess?

Illustration

Interactive Turn the parts of the request on and off.Try it on the web: aifromthepartsup.com/asking-well/1/

Each part you leave out is something the model has to guess.

Your request

  • The job Write a welcome email for three new hires on the accounting team.
  • Who it’s for They start Monday. Two are recent graduates; one is moving here from Kansas City.
  • Background Attached: the first-week schedule and the parking map.
  • An example Here is last year’s welcome email. Match its friendly tone.
  • Limits Under 150 words. Don’t mention pay or benefits.
  • The shape A subject line, then three short paragraphs.

What it has to guess: 0

  • What is this for? Without the job, it has to guess what you want made.
  • Who is reading? It writes for a generic reader, in a generic tone.
  • What actually happens that week? It fills the gap with details that sound right and may be wrong.
  • What does good look like here? It picks a style for you.
  • How long, and what is off limits? It may run long or include things you’d cut.
  • What shape should it take? You may get one long block to rework.

Nothing: every part is there.

An example request written for this page. No AI output is shown: the point is what you leave it to guess.

Go deeper: why examples work

Examples work because a model picks up a pattern from them on the spot. GPT-3 could do many tasks it wasn’t specially trained for from a few examples written into the prompt, with no retraining, though it still struggled on some.3

Notes (3)
  1. 1Anthropic, “Prompting best practices,” Claude Platform Docs, n.d., https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices (accessed October 27, 2026).
  2. 2OpenAI, “Prompt engineering,” OpenAI API Docs, n.d., https://developers.openai.com/api/docs/guides/prompt-engineering (accessed October 27, 2026).
  3. 3Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al., “Language Models are Few-Shot Learners” (preprint, arXiv, May 28, 2020), https://doi.org/10.48550/arXiv.2005.14165. Cited at abstract; p. 1.

Chapter 13 · page 1 of 2

Turn pages with the ← and → keys.

Cite this page

https://aifromthepartsup.com/asking-well/1/

Using this book

© 2026 Steve Sparks. All rights reserved. You're welcome to quote short passages and cite any page, with a link back. To reuse diagrams or larger parts, ask through GitHub.

Quoted material and figures from other sources remain their owners'.