Giving examples and a role to set the style

1Two examples beat ten adjectives

Language models learn in context: a few examples placed in the prompt are enough to make them reproduce a format, a phenomenon documented since 2020 by Brown and co-authors. Anthropic's guide recommends three to five examples to stabilise a format; for a simple in-house template, two are often enough. The technique is called prompting with examples, or few-shot.

Which examples? Real productions from your team, anonymised: three headlines already published on the intranet, a meeting report considered exemplary, a client reminder that worked. In Copilot Chat, you paste them into the prompt, or you attach the model document as a file, writing "reproduce this structure". Pair them with a checkable rule: "60 characters at most, no exclamation mark". The example shows, the rule bounds.

The role is more debated. A study of 162 roles, four model families and 2,410 factual questions shows that personas do not improve correctness, and that the best role for a given question is unpredictable. So the role serves tone and structure, never the reliability of figures.

Write it as a job function useful to the result: "you are the assistant to the head of purchasing", not "you are the world's best expert". And leave it out when the function changes nothing in the answer.

One trap of the technique: the example contaminates. If your three headlines are all about the staff restaurant, Copilot will propose headlines about catering. Pick examples that vary in topic and are identical in form. Finally, the terms of your trade must be kept as they are: ask for it, otherwise the model rewords them in generic language.

Tap to enlarge

2Ten intranet headlines built on three models

A communications officer wants ten headlines to announce the new bike shelter on the intranet.

Turn 1 (Framing with reference examples and bounds).

You are the intranet editor of an industrial company. Suggest ten headlines for an article announcing the new covered bike shelter, opening on 3 November. Expected style, see these three published headlines: "The staff restaurant goes organic", "Remote work: the new agreement in five questions", "Welcome to this autumn's 12 apprentices". Rules: 60 characters at most, no exclamation mark, a verb or a colon as in the examples. Numbered list.

Copilot outputs ten tailored suggestions between 38 and 58 characters, without exclamation marks, patterned after the models, such as "Bike shelter: 40 covered spaces from 3 November".

Turn 2 (Adaptation for digital workshop screens).

From these ten proposals, select the top three for digital workshop screens. Constraints: 35 characters maximum per line, infinitive or noun phrase for instant readability at 5 metres.

Observed Copilot response:

1. Bike shelter: 40 spaces from 03/11 (35 chars)
2. New covered bike shelter open (31 chars)
3. 40 covered bike spaces on Nov 3 (31 chars)

What changes. The examples set the form better than adjectives. The numbered rule lets you check every headline by counting. In Turn 2, Copilot applies the length and immediacy filter without altering the factual content.

3Give your progress update its style back

Every week your team writes a project status update whose style has drifted. Here is the initial, imprecise prompt:

"Write a clear, professional status update from my notes."

Improve it using this operational data from NovaTech Ltd:

"Raw Meeting Notes - NovaTech Ltd: 1) Karim validated the banking integration test suite with 0 blocking defects. 2) Sarah finishes the logistics user manual before Thursday 6pm. 3) David completed training for 24 out of 30 team leads, catch-up scheduled for June 28."

Construct your complete prompt following this structured pattern:

"Role: Project management assistant at NovaTech Ltd. Task: Draft the weekly status update from the raw notes. Expected style examples: - 'Billing module user acceptance sign-off completed without reservations (Karim, June 14)' - 'Barcode scanner hardware rollout live across Loading Dock B (Sarah, June 17)' Rule: Exactly one bullet per note (Action + Owner + Deadline), 15 words maximum per bullet, no conversational intro. Raw notes: [paste the notes above]."

Self-assessment grid: (a) the role represents a useful operational function; (b) two real examples bound the format; (c) every generated bullet rigorously respects the 15-word ceiling.

Open the prompt composer