Prompt Structures: Layout That Improves Results

Good prompts have visible structure: a clear task, separated context sections, examples, and a specified output format.

Clear and direct

The most important rule from Anthropic's prompting docs: be precise about what you want. The test: show your prompt to a colleague with no prior context. If they would be confused, the model will be too. Numbered steps help when order or completeness matters.

Separate sections — with XML tags

As soon as a prompt mixes several content types (instructions, context, examples, input data), Anthropic recommends XML tags: <instructions>, <context>, <example>. Each content type gets its own tag — so the model can tell unambiguously what is an instruction and what is just an example. Consistent, descriptive tag names matter.

Assign a role

Even a single sentence in the system prompt ("You are an experienced Python developer") noticeably focuses behavior and tone.

Show examples (few-shot)

Per the docs, examples are one of the most reliable ways to steer format and style. Recommended: 3 to 5 examples, relevant to your real use case and varied enough, each wrapped in <example> tags.

Specify the output format

Say what the model should do — not what it should avoid: "Respond in flowing prose" works better than "No markdown". For fixed structures, a format example or a dedicated tag for the answer helps.

EXAMPLE

<instructions> Summarize the text in 3 bullet points. Respond in English. </instructions> <context> Audience: beginners with no prior knowledge. </context> <text> [the text goes here] </text>

🛠️ EXERCISE — TRY IT YOURSELF

Take one of your own unstructured prompts and rebuild it with structure.

Pick a prompt you used recently that mixes instructions, context, and data.
Split it into sections with tags: <instructions>, <context>, and a tag for the input data. State the desired output format at the end.
Send both versions to your model and compare the answers for accuracy and format adherence.

✅ SELF-CHECK

☐ Can you clearly show in your prompt where the instruction ends and the data begins?
☐ Did you state the output format positively ('do X') instead of negatively ('no Y')?
☐ Would a colleague with no prior context understand your prompt on first read?

QUICK QUIZ

Why does Anthropic recommend XML tags like <instructions> and <context> in prompts?