How much context should a prompt include?

A good rule of thumb: give the model what a capable stranger would need to do the job well on the first try. Picture handing your request to a skilled freelancer who has never met you. If they would have to come back and ask "for whom?", "how long?" or "what's it for?", the prompt is missing context. If they would have to wade through details that change nothing, it has too much.

This idea sits at the heart of Chapter 4 of A Beginner's Guide to ChatGPT, Robert Scales's book published this spring. The chapter argues that context is the main lever you control. A model like GPT-4 builds its answer from the words you give it, so every relevant detail you add narrows the range of likely replies toward the one you actually want.

What counts as useful context

The book groups useful context into a handful of habits. Rewritten as questions you can ask yourself before you press Enter:

OpenAI's own guidance on prompting makes the same point from the engineer's side: be specific about context, outcome, length, format and style, and replace fuzzy wording with precise wording, such as asking for three to five sentences instead of "fairly short."

Notice that one row is about removing things. The book warns that a prompt crammed with excess detail or competing directions can lead the model astray. More context is better only while it stays relevant.

The canoe-trip example

The book's clearest illustration comes from Chapter 3, built around canoe camping, a pastime the author shares with his wife. The test is simple: ask for five evening activities at camp, once with almost no context and once with a full picture.

The bare version reads roughly like this:

Give me five things to do in the evening on a canoe camping trip.

The detailed version sets a role and describes the actual day. Here it is in fresh words that keep the same ingredients:

Act as a seasoned wilderness guide who knows the Canadian backcountry,
camp cooking and outdoor survival, and who likes creative solutions.

Two of us are canoe tripping. Today we paddled about 12 km and carried
roughly 80 kg of gear over a 2 to 3 km portage. Suggest five simple,
low-effort activities for the evening while we pitch the tent, cook,
and watch the sunset by the fire above the lake.

According to the book, the short request produced a sensible but generic list: campfire cooking, stargazing, watching wildlife, telling stories, an evening paddle. Nothing wrong, nothing personal. The detailed request produced suggestions shaped to two tired paddlers, such as storytelling games, simple fire cooking, gentle stretching by the shore and quiet nature art, plus a closing reminder about Leave No Trace and wildlife safety. Same model, same topic; the difference was what the prompt told it.

A quick way to check your prompt

Before sending, read your prompt and look for four things: the role or audience, the situation, the goal, and the format. If one is missing, add a sentence. Then look for anything that would not change the answer if you deleted it, and cut it. The book also stresses checking whatever comes back, because models can state invented facts with confidence. Context improves the fit of an answer; it does not guarantee accuracy.

What happens when you give too little?

The model fills the gaps with the most typical answer. That is the core issue. A language model works by predicting likely text, so when your prompt leaves the audience, purpose and format open, it falls back on what is common across everything it has seen. You get the average answer to the average version of your question.

In the canoe example, the short prompt said nothing about distance, fatigue, the number of people or the setting. So the model answered for "people camping" in general. Each suggestion was reasonable, but none reflected two people who had just hauled heavy packs over a portage. The reader then has to do the tailoring alone, which defeats much of the purpose.

Too little context also costs time. You end up in a loop of follow-ups: "shorter," "for beginners," "in Canadian terms." The book treats that loop as normal and useful, recommending that you refine a prompt step by step after reading each reply. But much of the loop can be skipped by putting those details in the first message.

The fix is rarely a long prompt. Usually two or three well-chosen sentences, naming who, what for and in what shape, move an answer from generic to useful.