S1
One Request: Instructions, Examples, Constraints
Build a single model request that is clear, testable, and no more complicated than necessary.
- 01Write instructions that define the task, boundaries, and audience
- 02Use examples when rules are easier to show than describe
- 03Turn vague quality words into observable checks
A practical prompt contract
worked exampleA useful request answers 5 questions:
- Task — What should the model do?
- Input — What material should it use?
- Boundaries — What must it avoid or refuse?
- Output — What should the response contain?
- Success test — How will you know it worked?
Replace adjectives with checks
“Be accurate, helpful, and concise” sounds clear but is hard to test.
Prefer:
Use only facts in the incident log. Write 3 bullets under 18 words each. Include customer impact, time to recovery, and the next action. If a fact is missing, write “unknown.”
Now you can check grounding, shape, length, and missing-data behavior.
Use examples to show a pattern
Examples are useful when the rule has exceptions or a domain-specific style.
Input
Service returned 503 for 7 minutes. 14% of requests failed. Cache restart restored traffic.
Output
- Checkout requests failed for 7 minutes.
- 14% of requests returned errors.
- Restarting the cache restored service.
The example teaches structure and tone without a page of abstract instructions.
Common MisconceptionA longer prompt is automatically more precise.
CorrectionLong prompts often repeat or conflict. Precision comes from observable constraints and representative examples.
Deep dive: Reasoning instructions
Reasoning prompts can improve some multi-step tasks, but they are not a universal upgrade. Evaluate the final task outcome, latency, and cost. When the reasoning itself is sensitive, ask for a concise justification or structured evidence rather than exposing unrestricted internal reasoning.
References & deeper reading (2)
- Prompt Engineering GuideOpenAI Platform
- Prompt Engineering OverviewAnthropic
Retrieval Practice
Check one idea at a time
Which instruction is easiest to evaluate?