Few-Shot Framework: Email Response Styles

Use this when you need consistent, tone-accurate email drafts from an LLM. The few-shot examples teach the model what each tone actually looks like in practice, rather than relying on vague instructions like "be professional" or "be casual."

This is particularly useful for teams. If multiple people use the same prompt, the few-shot examples ensure everyone gets the same interpretation of "formal" or "friendly." It removes the ambiguity that single-word tone instructions create.

Also useful for non-native English speakers who want to hit the right register in professional emails without overthinking it.

Few-shot examples beat adjectives. Telling a model "write in a friendly tone" is vague. The model's idea of "friendly" may not match yours. But showing the model a concrete example of what "friendly" looks like in your context gives it a precise target to mirror. This is the core principle of few-shot prompting: demonstration over description.

Same scenario, different tones. The three examples all respond to the same situation (rescheduling a meeting). This isolates tone as the only variable. The model learns: same content, different delivery. If each example used a different scenario, the model might conflate the tone difference with the content difference.

The examples are short and realistic. Long, elaborate examples teach the model to be long and elaborate. These examples are the length of real emails. The model mirrors not just the tone but also the length and structure.

The template variables ({{scenario}} and {{tone}}) create a clear contract. The model knows exactly what inputs to expect and what output to produce. No ambiguity about the task.

Add your own tones. Three is a starting point. Common additions: "empathetic" (for difficult news), "urgent" (for time-sensitive requests), "persuasive" (for sales follow-ups). For each new tone, write one example response to the same base scenario.

Replace the examples with your own writing. The most effective few-shot prompts use examples written by you, in your actual voice. Take three real emails you have sent, one per tone, and swap them in. The model then mirrors your specific style, not a generic approximation.

Add more example pairs. Two to three examples per tone gives better consistency than one. If you find the model drifting from your preferred style, add a second example for that tone using a different scenario.

Extend beyond email. This same framework works for Slack messages, LinkedIn messages, customer support replies, or any format where tone consistency matters. Change the scenario type and examples to match.

Few-shot prompts use more tokens than zero-shot instructions. Each example adds to the input length. For models with smaller context windows or if you are optimising for cost, keep examples concise.

The model may blend tones if the scenario is emotionally complex and the tone instruction is simple. For example, delivering bad news in a "direct" tone may come across as cold. For sensitive scenarios, consider adding a fourth tone like "direct but empathetic" with its own example.

Few-shot examples are most effective when they are close to the actual task. If your real emails are very different from the provided examples (different industry, different relationship type), the model may produce something that technically matches the tone but feels off for your context. Replace the examples with your own.

Claude: Excellent at mirroring few-shot examples. Tends to match tone, length, and structure very precisely. If anything, Claude may follow examples too literally, so ensure your examples represent the output you actually want.

GPT: Mirrors tone well but may add extra pleasantries not present in the examples (especially in "direct" tone). Reinforce with: "Match the length and structure of the examples. Do not add sentences beyond what the examples demonstrate."

Gemini: Works reliably with few-shot examples. May default to slightly longer output than the examples show. Keep examples concise to counteract this.

General: Few-shot prompting is one of the most universally effective techniques across all models. This framework transfers to any LLM without modification.