The AI Learning Hub Journal

Everything in the Box

A prompt is everything the model receives in one requestnot just the question you typed — up to four kinds of content travel in the same boxONE REQUEST — EVERYTHING THE MODEL WILL SEEINSTRUCTIONthe verb — summarise,draft, compare, check,rewriteCONTEXTwho it is for, what it isfor, and what mustbe trueEXAMPLESone or two samples ofan answer you wouldacceptDATAthe material itself,pasted in — not describedfrom memoryTHE WEAK PROMPT — INSTRUCTION ONLY"Write an email about the delay."context — missingexamples — missingdata — missingTHE STRONG PROMPT — ALL FOUR PARTSINSTRUCTION"write the delivery-delay email"CONTEXTour fault, second slip, keep the clientLIMITSdirect, apologetic, under 150 wordsDATAthe facts to use, pasted inLABEL THE PARTS — TASK FIRST, PASTED MATERIAL LAST UNDER A HEADINGthe model reads one continuous text with no slot marked "important" — a heading like "Draft to edit:" tells it which part is whichTHE MODEL ANSWERS FROM WHAT IS IN THE BOX — AND NOTHING ELSEthe test: could a new colleague do the task from your message alone? if not, add the missing parts
A prompt is everything sent in one request — strong prompts add the context, examples and data that weak prompts leave the model to guess.

Four Kinds of Content, One Request

Most people picture a prompt as the question they typed. It is broader than that. A prompt is everything the model receives in a single request, and it usually contains up to four kinds of content. The instruction is what you want done. The context is the background that makes the task make sense — who you are, what this is for, what constraints apply. The examples are samples of what a good answer looks like. The data is the raw material to work on: the email to reply to, the notes to summarise, the numbers to check. Weak prompts almost always contain only the instruction. Strong prompts contain the other three as well, because those are the parts that carry the information the model has no other way to get.

  • Instruction: the verb — summarise, draft, compare, check, rewrite
  • Context: who it is for, what it is for, what must be true
  • Examples: one or two samples of the output you would accept
  • Data: the actual material, pasted in, not described from memory

Weak and Strong, Side by Side

Weak: "Write an email about the delay." Strong: "Write an email to a client whose delivery has slipped by two weeks. Context: the delay is our fault, a supplier missed a deadline, and this is the second delay on this account. They are annoyed but we want to keep them. Tone: direct and apologetic, no corporate padding. Length: under 150 words. Do not offer a discount — I have not approved one. Here are the facts to use: [paste]." Both are one message. The second one takes forty seconds longer to write and produces something you can send rather than something you have to rewrite. Nothing clever is happening in the second version. It just contains the information a competent colleague would have needed too.

  • Weak: "Write an email about the delay" — no reader, no history, no limits
  • Strong: names the reader, the cause, the relationship, the tone, the length
  • The strong version also says what NOT to do, which prevents a whole class of bad drafts
  • Test: could a new colleague do this task from your prompt alone? If not, it is underspecified

Order and Labels Help More Than You Expect

The model reads your prompt as one continuous piece of text. It has no special slot marked "this is the important bit". So when a prompt mixes instructions, background and pasted data into one paragraph, the boundaries get blurry — and you occasionally get the model treating your data as instructions, or your instructions as something to summarise. The fix is boring and effective: label the sections. Put the task at the top, the constraints next, and the raw material at the bottom under an obvious heading such as "Document to review:". Frontier models generally handle long, well-structured prompts far better than long, unstructured ones. Structure is not decoration; it tells the model which part is which.

  • Label pasted material explicitly: "Transcript:", "Draft to edit:", "Data:"
  • Put the task first so everything after it is read as serving that task
  • Keep constraints together rather than sprinkled through the paragraph
  • Long is fine; long and undifferentiated is where things get lost

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.