Ready to put this into action?
Get the complete AI Integration Playbook β Practical AI implementation guide β prompt engineering, workflow automation, and ROI frameworks.
Article 013 Β· Part 2
Show an Example of What Good Looks Like
Use a small demonstration to make your expectations visible.
By Randy Salars Β· Published
On this page
- An example guides the current task
- Show the source record as well as the finished description
- Include a case with missing information
- Explain a counterexample
- Try the pattern on new items
- Check whether the example transferred correctly
- Students: copy the reasoning structure, not somebody's assignment
- Your exercise
Use a small demonstration to make your expectations visible.
Maya has three used items to describe for a neighborhood sale. She asks an assistant to write short descriptions.
One comes back as a cheerful paragraph. Another looks like a technical specification. The third declares an ordinary storage box βa must-have solution for every modern home.β
Maya wanted three consistent, factual entries. She could keep explaining what βconsistentβ means. Or she could show it.
A good example makes part of the task visible. It can demonstrate structure, tone, the amount of detail, and what to do when a fact is missing. The trick is to explain which parts to copy and which parts belong only to the example.
An example guides the current task
Suppose you paste a short description into a conversation and say, βUse this structure.β The assistant can use that material as context for its next answer. This is often called few-shot prompting when you supply a small number of demonstrations.
Supplying an example in a prompt does not itself retrain the model's underlying parameters. Whether a service separately retains content or uses it for future training is a different question governed by its settings and policies.
For the task at hand, think of the example as a sample on the workbench. It shows the form you want the next piece to take.
Anthropic's documentation recommends examples that are relevant, varied, and clearly separated from instructions. Those principles help prevent the assistant from learning the wrong pattern from your demonstration. Anthropic: Use examples effectively.
Show the source record as well as the finished description
A polished answer alone can be ambiguous. Which statements came from the source? Which were invented for style?
Maya supplies a fictional record and its matching description:
Example input
- Item: Ceramic planter.
- Height: 8 inches.
- Condition: Small chip on rim.
- Drainage hole: Present.
- Price: $6.
Example output
Ceramic planter β $6
Eight inches tall, with a drainage hole. A small chip is present on the rim.
Then she explains the pattern:
Start with the item name and supplied price. Follow with one or two plain sentences covering the supplied features and condition. State defects directly. Do not add claims about quality, rarity, brands, or suitability that are absent from the record.
Now the assistant can distinguish the structure from the planter's particular facts.
Include a case with missing information
One tidy example may teach the assistant that every item has a price, size, and known condition. Real records are often incomplete.
Maya adds a second demonstration:
Example input
- Item: Desk organizer.
- Compartments: Three.
- Condition: Scratches on base.
- Price: Not supplied.
Example output
Desk organizer β price not supplied
Three compartments. Scratches are present on the base.
This shows that an unknown field should remain unknown. The assistant should not borrow the planter's $6 price to make the entries look complete.
Maya could omit price from all descriptions and list missing fields separately. Either approach can work. What matters is choosing a consistent policy before producing the batch.
Explain a counterexample
A counterexample shows what fails and why. It is most useful when it points to a likely misunderstanding.
For the planter, this would be unacceptable:
A flawless artisan planter that will transform any room. A rare find at only $6!
Maya labels the problems: βflawlessβ contradicts the chip; βartisanβ and βrareβ lack support; βtransform any roomβ adds promotional noise.
The lesson is not that every lively adjective is forbidden. It is that the description must preserve the source record and serve the reader's decision.
Always mark a counterexample clearly. A pile of unlabeled good and bad outputs can create conflicting instructions.
Try the pattern on new items
Here are three fresh fictional records:
| Item | Supplied features | Supplied condition | Price |
|---|---|---|---|
| Canvas tote | Two handles; inside pocket | Small ink mark inside | $4 |
| Table lamp | Height 16 inches | Switch works in the recorded check; shade has a dent | $9 |
| Metal bookends | Pair | Scratches on both bases | Not supplied |
Maya's request is:
Describe these three items using the demonstrated format. Use each item's own facts only. If price is missing, say βprice not supplied.β Do not transfer features from one item to another. Report any conflicting input instead of selecting a fact silently.
An acceptable set would be:
Canvas tote β $4
Two handles and an inside pocket. A small ink mark is present inside.Table lamp β $9
Sixteen inches tall. The switch worked in the recorded check; the shade has a dent.Metal bookends β price not supplied
A pair of metal bookends. Both bases have scratches.
Notice what the lamp description does not establish: full electrical safety, remaining life, or a guarantee. The supplied switch check supports only the narrower statement.
Check whether the example transferred correctly
Consistency is only one part of quality. Three equally polished descriptions can share the same invented claim.
Compare each description with its own record. Check identity, numbers, condition, and unknowns. Then look for facts that leaked from an example or neighboring item.
For this exercise, a drainage hole in the tote description is an obvious leak. More subtle leaks include a copied price, a missing defect, or the word βpairβ applied to a single item.
If a failure occurs, repair the rule or add one useful example addressing it. Do not keep adding demonstrations merely to make the prompt look sophisticated. A small, coherent set is easier to maintain.
Students: copy the reasoning structure, not somebody's assignment
A student can use an authorized sample to understand how an explanation is organized. For example, a science teacher's model paragraph might separate an observation, an interpretation, and a limitation.
A learning request could be:
Identify the purpose of each sentence in this teacher-provided example. Then give me a blank structure to use with my own observations. Do not transfer the example's results into my report.
The student fills the structure with their actual evidence and follows the assignment's AI rules.
Be especially careful with complete student essays or reports. Another student's work is not a reusable fact bank. Use material you are authorized to use, and preserve required attribution. A structural demonstration should help you understand the work, not conceal who did it.
Your exercise
Create your own descriptions for the three records above. Add one new fictional item that has a missing field and one that contains conflicting prices.
For the missing field, apply the stated policy. For the conflicting prices, require a question or a flagged conflict rather than a guessed answer.
Keep the source records, example set, and resulting descriptions together. That lets you see whether the instructions still work when you reuse them later.
Completion check: The new descriptions follow the intended structure, preserve defects and unknowns, and import no facts from the examples.
Get the AI Dispatch
Weekly insights on ai & technology β delivered to your inbox. No spam, unsubscribe any time.
Want to choose specific topics? Customize your interests