Free mini-course · Lesson 4 / 6
Use examples to guide AI
When the result must follow a style or a rule that is hard to describe, show one or two examples. This is the idea behind few-shot prompting: guide the answer with examples in context.
Editorial team: Take Two ·
The idea
Choose examples consistent with the rules and include an ambiguous case. Examples demonstrate the desired behaviour within the request; they do not permanently train the model. For a simple task, try instructions alone first and add examples when they address an observed problem.
A vague request
Classify these requests by priority.
Why it is incomplete
Priority depends on context. An urgently worded message does not establish that a service is down.
A prompt to try
Try the prompt in your usual AI tool. Select and copy the text; no coding is needed. Access, features and limits depend on the tool and plan you use.
Classify each request as: DOWN, DEGRADED, INFORMATION, CLARIFY.
Rules: DOWN = unavailable service; DEGRADED = available service with problems; INFORMATION = instructions requested with no failure described; CLARIFY = insufficient facts.
Do not rely on capital letters or tone.
Format: category — short reason supported by the text.
Examples:
“The portal will not open for anyone.” → DOWN — unavailable for everyone.
“The portal works but each page takes 40 seconds.” → DEGRADED — available but slow.
“How do I change my password?” → INFORMATION — instructions requested.
“URGENT, I have a problem!” → CLARIFY — no description.
Requests to classify:
1. I cannot log in; my colleagues are working normally.
2. Where is the manual?
3. After the update, the service is slow but reachable.What to look for
The first request needs clarification: it could be an account problem or unavailability for that person. The other two are INFORMATION and DEGRADED respectively.
Now try it
Your exercise
Add “The service became available again after five minutes” and “I want to activate a new account”. Which categories are missing? Propose revised rules, then retry all cases.
Check the result
- Do the reasons cite facts in the message?
- Are ambiguous cases flagged for clarification?
- Do the categories cover the new requests?
See the commented answer
The original categories do not distinguish a resolved incident from a service request. Add RESOLVED and SERVICE_REQUEST with consistent definitions and examples. Do not change a rule just to pass one case while ignoring the others.
Examples and commented answers are teaching material written for this course. They are not recorded model outputs; answers may vary.
Sources and further reading
The exercises are original Take Two teaching examples. For technical background on instructions, examples and evaluation criteria, see: