Few-Shot Prompting

When you provide examples in your prompt, the model uses in-context learning to identify the pattern and apply it to new inputs. The model is not being trained—its weights do not change. Instead, the examples act as demonstrations that prime the model to produce output in the same format and style. For coding tasks, this means you can show the model one or two examples of how you want a function structured, and it will follow that pattern for subsequent generations.

Few-shot is particularly powerful for enforcing coding conventions that are hard to describe in words. Instead of explaining your team's error handling pattern in abstract terms, show two examples of functions that follow the pattern. The model picks up on naming conventions, error structures, logging formats, and architectural patterns from examples much more reliably than from descriptions.

// Few-shot prompt for consistent API route handlers
// Prompt:
// "Generate a PATCH /users/:id endpoint following this pattern:
//
// Example 1:
// app.get('/products/:id', auth, async (req, res) => {
//   try {
//     const product = await Product.findById(req.params.id);
//     if (!product) return res.status(404).json({ error: 'Product not found' });
//     res.json({ data: product });
//   } catch (err) {
//     logger.error('GET /products/:id', err);
//     res.status(500).json({ error: 'Internal server error' });
//   }
// });
//
// Example 2:
// app.delete('/products/:id', auth, async (req, res) => {
//   try {
//     const product = await Product.findByIdAndDelete(req.params.id);
//     if (!product) return res.status(404).json({ error: 'Product not found' });
//     res.status(204).send();
//   } catch (err) {
//     logger.error('DELETE /products/:id', err);
//     res.status(500).json({ error: 'Internal server error' });
//   }
// });
//
// Now generate PATCH /users/:id:"

In Claude Code, the most effective form of few-shot prompting is referencing existing files: "Follow the pattern in routes/products.ts." The agent reads the file and applies the same patterns, which is even better than pasting examples into your prompt.