Give the model what it cannot know
Supply the facts, material and situation the task depends on, clearly separated from your instructions.
هذا الدرس غير متاح بلغتك بعد، لذا يُعرض بلغة أخرى (English).
Why it matters
A model knows a great deal about the world and nothing about your Tuesday meeting. Every prompt rests on facts that live only in your head, your files or your organisation: the notes, the data, the history, the reason this is happening now. If they are not in the prompt, the model fills the gap with plausible guesses, and plausible guesses are the hardest errors to spot.
Anthropic's guidance treats context as a lever on quality: explaining the situation and the motivation behind a request gives the model more to work with and produces more targeted answers. Its article on context engineering goes further and treats prompting as curating every piece of information the model sees. OpenAI describes context as the information the model cannot otherwise reach.
Context also needs edges. When pasted material runs straight into your instructions, the model cannot always tell which sentences are data and which are requests. Tags such as <notes>…</notes>, fenced blocks or clear headings settle that.
What good looks like
- 0 — No context. The model must guess the subject, the situation and the material: “Summarise this.”
- 1 — Situation named in passing. “Summarise the meeting we had about the budget.” The meeting itself is missing.
- 2 — Material present but undelimited or incomplete. The notes are pasted, but they run into the instructions with no boundary.
- 3 — Material delimited and situation explained. Notes in tags, plus a line on what the meeting was and who the summary is for. Small gaps remain.
- 4 — Complete, delimited, with what is missing stated. Material, background the model could not infer (“the third review this quarter; last time travel was overspent”), and a note on what it does not have (“June figures are not available, so do not estimate them”).
Common mistakes
- Referring to material instead of including it (“the report”, “our policy”, “last week's email”).
- Pasting everything you have. More text is not more context; include what the task needs and say why it is there.
- Leaving pasted material undelimited, so an instruction buried inside it may be followed by accident.
- Omitting the background that explains why the task exists, which is often what decides the answer.
- Forgetting to say what is unknown, which invites the model to invent it.
Try it on your own prompt
Paste a prompt you rely on into the workspace and read it as if you were a capable stranger with no access to your files. Which fact would you have to ask for first? Add it, wrap any pasted material in tags, and add one line on what is not available. Then compare the context score before and after.
أمثلة محلولة
Meeting summary
Summarise the meeting we had about the new opening hours.
<notes>
{{paste the meeting notes here}}
</notes>
These are my notes from Tuesday's staff meeting at our community library about opening until 8 pm on Thursdays.
Background you could not know from the notes: this is the second meeting on the subject. The first ended without agreement because the volunteer rota was not ready. The council has not yet decided whether to fund the extra hours, so leave the budget out rather than speculating.
Summarise the notes for the volunteers who could not attend: what was agreed, what is still open, and who is doing what by when.
The weak prompt refers to a meeting the model never saw. The strong one supplies the notes inside tags, adds the history the notes leave out and states what is not yet known.
Debugging a script
Why is my code broken?
<code language="python">
{{paste the function here}}
</code>
<error>
{{paste the full error message and traceback here}}
</error>
Situation: this function reads the CSV export from our booking system and totals deposits by month. It worked until last week, when the export started including a currency symbol in the amount column ("£45.00" instead of "45.00"). I am on Python 3.12 with pandas 2.2 and have not changed the function. I no longer have the old export files, so work from the description above rather than asking me to compare them.
Explain what is causing the error and propose the smallest change that fixes it.
Without the code, the error and what changed, any answer is a guess. The strong version delimits each input, explains the situation and says which information is unavailable.
تمرين
Rewrite this prompt so the model knows what the material is, where it starts and ends, and what counts as important.
Rewrite this prompt with the facts the model needs to judge the price, and mark clearly what you do not know.
Rewrite this prompt so the essay and the teacher's feedback are included (or left as placeholders), delimited, with the situation explained.