An example beats a description
Try describing the style of your Telegram channel in words. You'll get something like "lively but to the point, with light irony" — a thousand different styles fit that description. Now just show three of your posts — and the model catches the style more precisely than any description.
The technique is called few-shot prompting: instead of explaining, you show samples. According to Anthropic, examples are the most underused prompting tool and at the same time one of the most powerful.
How to build a prompt with examples
The structure: task → examples → new input.
Rewrite dry notifications in the style of our service.
Example 1: Before: "Your subscription expires on 02/01" After: "Your subscription is living out its last days — renew while everything still works"
Example 2: Before: "Payment failed" After: "The bank blocked the payment. It happens. Try another card"
Now rewrite: "Account blocked for violating the rules"
Rules for working with examples:
- Two or three examples are usually enough; one is already better than none.
- Variety matters more than quantity. The examples should cover different cases, not repeat one.
- Quality is critical. The model will copy the flaws of the samples too — a botched example teaches botched output.
- Edge cases — show how to handle a non-standard input, if one is possible.
Examples convey beautifully what's hard to put into words: tone, rhythm, degree of formality, document structure. A report format is easier to show with one sample than to describe in a paragraph.
Roles: the model's point of view
A role at the start of a prompt ("You are an auditor," "You are a picky editor") tunes three things:
- Criteria. An auditor looks for discrepancies, a marketer for growth points. They read the same document differently.
- Depth. "Explain like a physics teacher to a fifth-grader" and "explain like a physicist to a colleague" give different levels of detail.
- Tone. A lawyer writes differently from a blogger.
A role doesn't add knowledge to the model — it chooses which of the available knowledge to use, and in what manner.
The most underrated kind of role
"You are a skeptical critic. Tear this plan apart: find the weak spots, the risks, and the unfounded assumptions."
By default the model is tuned to be friendly: it praises, agrees, softens. For real decisions friendliness is harmful — you need a sparring partner, not a support group. The critic role removes that tuning. A combo move: first ask the model to improve a text, then, in the critic role, to attack the improved version.
Role + examples together
The maximum effect comes from the pairing:
You are the editor of our blog. Here are two samples of texts in our style: [examples]. Edit the article below to match this style. Whatever's off in the article's facts or logic — flag it with a comment [question: ...] right in the text.
The role sets the position, the examples set the specifics of style, the instruction sets the format for handling problem spots.
Key takeaways
- An example conveys style and format more precisely than any verbal description.
- Two or three varied, quality examples; the model copies the flaws of the samples too.
- A role tunes criteria, depth, and tone, but doesn't add knowledge.
- The critic role removes the default friendliness — a mandatory technique for checking decisions.