1.3Give examples (few-shot)
The essentials in 30 seconds
- Showing beats describing: one example of the expected result is worth ten lines of explanations — structure, tone and length are mirrored.
- Invest in your benchmarks: ONE perfect hand-written specimen is reused for every series (fact sheets, standard replies).
- Anonymise your examples: the model only needs the FORM — never real data on an unapproved tool.
Threaded mini-case — Amine produces his series of fact sheets in Algiers
Amine must write progress sheets for five sites on the same model and first receives five mismatched texts. He hand-writes ONE perfect sheet (the benchmark), then prompts: “Here is the model: [his sheet]. Write the four other sites on exactly this model”. He gets a consistent series in one shot. For his emails, he pastes ten lines of his prose and asks “in the style of this text”: his messages keep his voice.
Sequel to cases 1.1-1.2 — Amine shows instead of describing.
1. Showing beats describing
The most powerful and least used technique: show an example of the expected result. “Write the 4 others on this model: [your example]” — and the AI reproduces structure, tone, length and style with a fidelity no description reaches.
This is few-shot prompting: one or two examples are worth ten lines of explanations. Amine verifies it on his site sheets: describing the format had failed, the example succeeds first time.
2. Series and style transfer
Choose a good example (representative, exact format), use it for series tasks (product sheets, standard replies, posts).
Its powerful variant: “in the style of this text” — having the AI write like you, from a sample of your prose. Amine's texts keep his voice instead of the recognisable smooth style.
3. The right example, without real data
One good example is often enough; two or three help when the task has varied cases — beyond that, returns diminish. And if the tool is not company-approved: anonymise the example (names, fictitious figures) — the model only needs the FORM.
Amine replaces client names and amounts with fictitious values in his benchmark: the structure to imitate stays identical, the risk disappears (see 5.1).
🔎 Concrete example — 🔎 Concrete example — A series in one prompt
Few-shot + style transfer. Test with your own anonymised benchmark.
“Here is the model: [your perfect sheet, fictitious data]. Write on exactly this model the following items: [list of 4 topics].”
→ structure, length and tone mirrored.
“Write the following announcement in the style of this text: [10 lines of your prose]” → your voice.Reading: the AI imitates the demonstrated format with a fidelity no description reaches.
Lesson: invest in reusable “benchmark examples” for unlimited reuse.
Section vocabulary
- Few-shot
- Providing one or more examples of the expected result in the prompt: the model imitates the demonstrated format.
Simple example: The benchmark sheet Amine pastes before requesting the four others.
What it is not: describing a format at length without ever showing it.
Reflex: always show for a precise format.
- Zero-shot
- Asking without an example (the default mode): fine for simple tasks, risky for precise formats.
Simple example: Amine's simple request (“summarise this paragraph”) with no attached example.
What it is not: few-shot with a benchmark for a standardised sheet.
Reflex: zero-shot for the simple, few-shot for the formatted.
- Benchmark example
- YOUR best specimen of the task: the investment reused endlessly.
Simple example: The perfect sheet hand-written by Amine and kept in his library.
What it is not: a mediocre example stuffed with real data.
Reflex: one anonymised benchmark per series task.
- Style transfer
- “Write in the style of this text”: tone, rhythm and vocabulary imitated — your texts keep your voice.
Simple example: Amine's ten lines of prose attached to keep his tone.
What it is not: the default smooth style, recognisable a mile away.
Reflex: attach a sample of your writing for visible documents.
In practice — In practice — Produce a consistent series in one prompt
- Choose a series task from your work (sheets, descriptions, standard replies) and hand-write the perfect specimen of ONE item — with fictitious data if the tool is not approved.
- Prompt: “Here is the model: [your specimen]. Write on exactly this model the following items: [list of 4-5 topics]”.
- Check consistency: structure, length and tone must be mirrored.
- Style test: paste 10 lines of one of your emails and ask “Write [new request] in the style of this text” — do you recognise your voice?
Ready-to-say script — convince your team to adopt the benchmark (30 s)
“Stop describing the format: write ONE perfect sheet together, with fictitious data, and keep it as the benchmark. Then a single prompt — here is the model, write the others on it — and the whole series comes out consistent. We reuse the benchmark every time.”
Say it in a team meeting; co-write the benchmark on screen with Amine.
Copy-paste model — reusable few-shot prompt
Here is the model (BENCHMARK EXAMPLE, fictitious data):
---
[PASTE YOUR PERFECT SPECIMEN HERE]
---
Write on EXACTLY this model the following items:
1. _______________ 2. _______________ 3. _______________ 4. _______________
Constraints: same structure [ ] | same length [ ] | same tone [ ]
STYLE (optional): “in the style of this text: [10 lines of your prose]”First series
Start with ONE benchmark and four items: check consistency before scaling up — the benchmark's quality decides everything.
Amine's team in Algiers
Amine stores validated benchmarks (sheets, standard replies) in a shared folder: the whole team produces consistent documents, even newcomers.
Task with varied cases
If the task has two or three distinct cases: provide one benchmark per case rather than a single one — two or three targeted examples beat a block of explanations.
Sensitive example
Client data or real figures in the benchmark? Replace with fictitious values before prompting on an unapproved tool — the model only learns the FORM (see 5.1).
✓ Best practices
- Show, don't describe.
- One benchmark per series.
- Anonymise examples.
✗ Common pitfalls
- Describing without showing.
- Too many stacked examples.
- Real data not anonymised.
Key takeaways
- One example beats ten lines of description: show the expected result.
- Invest in your “benchmark examples”: they are reused for every series.
- Style transfer keeps your voice in generated texts.
- For strict formats (tables, standardised sheets), few-shot is almost mandatory.
48h challenge — apply tomorrow
Write ONE benchmark (fictitious if needed) for a series task and produce 4 items on it in one prompt (model above). 20 minutes is enough.
- I can explain few-shot vs zero-shot.
- I know how to anonymise an example (form only).
- I produced a consistent series + tested style transfer.
Frequently asked questions
How many examples to provide: one, two, ten?
One good example is often enough; two or three help when the task has varied cases. Beyond that, returns diminish and the prompt gets heavier.
My example contains real data, is that a problem?
Yes if the tool is not company-approved: anonymise the example (names, fictitious figures) — the model only needs the FORM. Reflex detailed in level 5.
What is few-shot prompting?