Module 5 Academy

Part 2 · Control

05

Few-shot prompting

3 min read · Module 5 of 13

The concept

Showing beats telling. You can spend a paragraph describing your house tone of voice and still get something off-key — or you can paste two examples of copy you've actually approved and the model will match the pattern with uncanny accuracy. This is called few-shot prompting: giving the model a small number of input→output examples so it infers the rule you couldn't articulate. It works because pattern-matching is precisely what the model does best, and because most of what makes your voice yours is tacit — rhythm, sentence length, how you handle the call to action — things you'd struggle to write down but can instantly recognise.

How many? Two to five. One example is often read as a template to copy too literally. Beyond five or six you get diminishing returns and you're burning context. Three is the sweet spot for tone; five if the task has variation you want covered.

How to pick them? Three rules:

  1. Pick your best work, not your most typical. The model regresses toward the average of what you show it. Show it your ceiling.
  2. Pick examples that differ from each other along the dimension that doesn't matter, and are consistent along the dimension that does. If you want tone consistency, show three different products in the same voice — that teaches "voice is the constant." If all three examples are for juice brands, it will think the task is about juice.
  3. Include a counter-example when the failure mode is predictable. "Here are two headlines we approved, and one we rejected — rejected because it led with the feature instead of the tension." Negative examples are extremely efficient.

Format them cleanly so the boundaries are unambiguous — labelled blocks, or INPUT: ... OUTPUT: ... pairs. Ambiguity about where an example ends is a real failure source.

Before → After

BEFORE 
Write Instagram captions in our brand voice. Our voice is friendly,
modern and a bit witty.

What you get: emoji-heavy, exclamation-marked, generic-agency captions that sound like every other brand. "Friendly, modern, witty" describes roughly 80% of all brands on earth.

AFTER 
Write 6 Instagram captions for [client]'s new product launch.

Match the voice in these three approved examples exactly — sentence
rhythm, punctuation habits, how the CTA lands:

EXAMPLE 1 (approved, café client)
"We open at 7. The croissants are gone by 9. This is not a threat,
it's a schedule."

EXAMPLE 2 (approved, gym client)
"Nobody's first session looks good. That's the entire point of a
first session. Tuesday, 6pm, no judgement."

EXAMPLE 3 (approved, optician)
"You've been squinting at your phone for four years. Come on.
Twenty minutes, one test, we'll be kind about it."

REJECTED EXAMPLE (do not write like this)
"🔥 Ready to upgrade your lifestyle? ✨ Our amazing new collection is
HERE and you won't believe the quality! Link in bio! 💯"
Rejected because: emoji clutter, hype adjectives, no specific detail,
CTA is generic.

Pattern to preserve: 2–3 short sentences. A concrete specific detail.
Dry humour. CTA is a plain fact, not a command. No emojis. No
exclamation marks.

New brief: [paste product brief here]
Exercise

Do this

Open your archive and pull the three best pieces of copy your agency has ever shipped for one client — the ones that survived every round of feedback intact. Paste them into a chat with nothing but: "Describe the voice rules these three share, as 8 specific rules a new copywriter could follow. Be concrete about sentence length, punctuation, and how the CTA works." You now have your brand voice written down — derived from evidence rather than from a brand book nobody reads. Keep it; you'll use it in Module 8.