What is few-shot prompting?

What is few-shot prompting and what changes in your prompt?

The mechanism is plain. You write two to a dozen worked examples above the real question and the model completes the pattern. Nothing about the model changes, because there is no training run, no adapter and no gradient. What changes is the distribution the next token is sampled from, because the examples now dominate the visible prompt and the model has been trained on data where a series of paired questions and answers is followed by another answer in the same shape.

The GPT-3 paper gave the technique its name and its taxonomy. Zero-shot is a task description with no examples, one-shot is a task description plus one example, few-shot is a task description plus several examples. Anthropic uses multishot for the same idea when it wants to fold one-shot into the label. All three run at inference time, all three cost only prompt tokens, and all three leave the model file untouched.

How is few-shot prompting different from fine-tuning?

Fine-tuning writes new weights. Few-shot prompting resends examples on every call. The consequences differ in every direction that matters to a shipping system.

• Cost per call. Few-shot pays for the example tokens on every request; fine-tuning pays once for the training job and then bills at the base per-token rate.
• Iteration speed. Editing a few-shot prompt is a text change. Editing a fine-tuned model is a new training run and a redeploy.
• What travels. A few-shot prompt is portable across providers and base models. A fine-tuned model is tied to the provider and base checkpoint you trained on.
• Where the behaviour lives. Few-shot behaviour is visible in the prompt, so a reviewer can read it. Fine-tuned behaviour is baked into weights and is only visible through evals.

The rule of thumb from every provider guide is to start with prompting, add examples when instructions alone are ambiguous, and reach for fine-tuning only when the prompt is long, hot and stable enough that paying the training cost beats paying the token cost of the examples on every call.

How many examples, and where do they go?

Anthropic's guidance is three to five diverse examples for most cases, with the caveat that quality and coverage matter more than count. The examples should include the edge cases you care about, use the same input format your production traffic uses, and show the exact output shape you want back. OpenAI's prompt engineering guide gives the same shape of advice and adds one concrete tip: use a stable delimiter to separate the demonstrations from the real query, so the model can tell where the examples end.

Placement matters because attention is not uniform across a long input. Put the examples close to the instruction and the real query at the end. If you are also carrying a system prompt and a tool schema, remember they share the same context window: a bank of twenty examples can push the system prompt out of the model's attention before the reply even starts.

Where few-shot prompting quietly breaks

Three failure modes come up often enough to plan for. The first is example bias: whatever pattern sits in your examples is the pattern the model copies, including biases you did not notice. If all five examples answer in one sentence, a case that needs three sentences gets one sentence. If every example uses one label from a set of ten, the other nine labels get picked less often at inference.

The second is contamination between examples and the query. If your examples happen to include text that resembles the real input, the model can answer from the demonstration instead of reading the query. This is easy to trigger with retrieval systems that draft examples from the same corpus the question came from.

The third is silent drift when you swap the model. Few-shot behaviour depends on how the model was trained to attend to demonstrations, and that changes across families and versions. When you upgrade the model, re-run your eval before you assume the same examples still work.

POST /v1/messages
{
  "model": "claude-sonnet-4-5",
  "max_tokens": 32,
  "system": "You label support tickets. Reply with one label from the set: billing, auth, shipping, other.",
  "messages": [
    {"role": "user",      "content": "Ticket: 'my card was charged twice this month'"},
    {"role": "assistant", "content": "billing"},
    {"role": "user",      "content": "Ticket: 'the reset email never arrived'"},
    {"role": "assistant", "content": "auth"},
    {"role": "user",      "content": "Ticket: 'parcel says delivered but I never got it'"},
    {"role": "assistant", "content": "shipping"},
    {"role": "user",      "content": "Ticket: 'how do I change the address on my invoice?'"}
  ]
}

response.content[0].text:
"billing"