- Practice
- About 16 min
- 5 practice questions
Context and instructions: background, audience, constraints, source text and iteration
Give an AI the background, audience, constraints and source text it needs, understand system versus user instructions, split long tasks into steps, invite "I don't know" and improve a prompt by iterating.
If something looks wrong, please tell us.
By the end of this lesson you can
- Add background, audience, constraints and source text so the answer fits your real need.
- Explain the difference between system instructions and user instructions, and split a long task into steps.
- Ask the AI to say "I don't know" instead of guessing, and improve a prompt over several rounds.
The previous lesson built the skeleton of a prompt. This lesson adds what turns a correct but generic answer into one that fits your situation: who you are, who the answer is for, what limits apply and what material to work from. Most disappointing AI answers come from missing context, not from a weak model.
Four kinds of context
- Background What you already know - the answer starts at the right point
- Audience Who it is for - vocabulary, examples and language suit the reader
- Constraints Your limits - length and style stay within them
- Source text Material you trust - the answer is tied to it, not the model's memory
Four kinds of context
| Context | Example line | What it changes |
|---|---|---|
| Background | I already understand simple interest and want to learn compound interest. | The answer starts at the right point and skips what you know. |
| Audience | Explain it for Class 8 students in a Hindi-medium school. | Vocabulary, examples and language suit the reader. |
| Constraints | Under 150 words, amounts in rupees, no technical jargon. | Length and style stay within your limits. |
| Source text | Use only the passage below. If it does not answer the question, say so. | The answer is tied to material you trust instead of the model's memory. |
Working from source text - a textbook paragraph, a public notice - is one of the strongest habits you can build. It reduces made-up content, but does not remove it, so you still compare the answer with the original.
System instructions and user instructions
Apps built on a language model usually give it system instructions (often called a system prompt): standing instructions written by the app's makers that apply to the whole conversation, such as "You are a tutor; give hints rather than full solutions". Your messages are user instructions and arrive within that frame. Users usually cannot see the system instructions, and they are generally meant to take priority when the two conflict. In your own chats you can create a standing instruction yourself: "For this whole conversation, answer in simple Hindi and keep each reply under 100 words."
Splitting a long task into steps
Splitting a long task
- Ask for an outline and check it
- One section at a time Each with its own instructions
- Review each section Correct errors early
- Combine the sections Ask for a combined version, or combine it yourself
- Ask for an outline first and check it.
- Ask for one section at a time, with its own instructions.
- Review each section before moving on; correct errors early.
- Finally ask for a combined version, or combine it yourself.
Shorter steps keep each request well inside the context window and make every part easier to check.
Letting the AI say "I don't know"
A model tends to produce some answer even when it has no good basis for one. Add: "If you are not sure, or the answer is not in the text, say 'I don't know' rather than guessing." This usually reduces invented answers. It is not a guarantee, so checking still matters.
Worked example 1: summarising a public notice for an office
A small panchayat office wants a staff summary of a public notice already displayed on its notice board, containing no personal details. Prompt: "Summarise the notice below for office staff in five bullet points. Use only the notice. If a date, time or place is not given, write 'not stated'. Do not add anything else." The reply lists the camp date as "not stated" - correct, because the notice gave only the month. Before circulating, the clerk compares every point with the original notice. Confidential files and documents with citizens' details would never be pasted into a chatbot.
Worked example 2: improving a prompt in rounds
| Round | Prompt | What was wrong with the reply |
|---|---|---|
| 1 | Explain inflation. | 400 words, unfamiliar terms. |
| 2 | Explain inflation in about 100 words for a beginner. | Better, but in English; the student studies in Hindi. |
| 3 | Now give it in simple Hindi as a table: Term, Meaning, Everyday example. | Usable. Student checks terms against the Hindi-medium textbook. |
| 4 | Ask me three questions to test whether I understood. | Turns reading into practice. |
Each follow-up names exactly what to change, and the earlier conversation stays in view, so there is no need to start again.
Context = background + audience + constraints + source text. Give standing instructions once, split long jobs into checked steps, invite "I don't know", and refine with specific follow-ups.
Key takeaways: tell the AI what you already know, who the answer is for, the limits it must keep and the text it should use. System instructions frame the whole conversation and normally outrank user requests. Break long tasks into checked steps, allow the model to admit uncertainty, and iterate with precise follow-ups - then compare the result with the source.
Check your understanding
-
1. Which line in a prompt adds a constraint?
- Keep the answer under 100 words and use amounts in rupees
- You are a commerce teacher
- Explain the meaning of GST
- I already know what a tax is
Show answer
Answer: A. A constraint sets limits on the answer, such as length and units. The role line says who to act as, the GST line is the task, and "I already know what a tax is" is background.
-
2. A student pastes a textbook paragraph and writes "Answer only from this passage." The reply includes a figure that does not appear in the passage. What is the best next step?
- Accept the figure, since the AI knows more than the passage
- Ask it to quote the sentence in the passage that supports the figure, say 'not in the passage' if there is none, and check the passage yourself
- Delete the passage and ask the question again without it
- Raise the temperature and ask again
Show answer
Answer: B. Asking for the supporting sentence exposes whether the figure came from the source. Checking the passage yourself settles it. Removing the source or adding randomness makes invented content more likely, not less.
-
3. Which description best fits system instructions in an AI app?
- The last message a user typed
- A setting that controls how random replies are
- Standing instructions from the app's makers that frame the whole conversation and are usually not visible to the user
- A list of the user's previous chats
Show answer
Answer: C. System instructions are written by those who build or run the app and apply to every reply. Users' messages are user instructions. Randomness is the temperature, and chat history is not a system instruction.
-
4. An officer needs a one-page brief of a 20-page report that the department has already published on its public website. Which approach is most sensible?
- Paste all 20 pages with "summarise" and circulate whatever comes back
- Ask for the one-page brief without giving the report
- Ask the AI to guess the report's content from its title
- Work in steps - outline, then each section's key points, then a combined brief - and check each part against the report
Show answer
Answer: D. Splitting the task keeps each step manageable and checkable, and comparing with the report catches errors. Circulating an unchecked summary, or asking without the source, invites mistakes and invented content. (A confidential report would not be pasted into an outside chatbot at all.)
-
5. Why add "If you are not sure, say 'I don't know' rather than guessing" to a prompt?
- It guarantees that every answer is correct
- It usually reduces invented answers, though the output must still be checked
- It makes the model search official websites
- It enlarges the context window
Show answer
Answer: B. Permission to admit uncertainty tends to cut down guessing, but it does not guarantee accuracy. The line does not trigger any web search or change the context window.
Sign in to keep track of the lessons you finish.