Claude (Anthropic) Prompting Guide: How Claude Thinks Differently

Claude is Anthropic's family of models — and while it can do most of what ChatGPT does, it has its own personality and rewards a slightly different prompting style. This guide shows you what makes Claude unique, when to reach for it, and the exact patterns that get the best results out of it.

Note: Claude's features evolve quickly. Check the official Anthropic documentation for the most current model names and capabilities.

1. Introduction

Claude was built with a strong emphasis on being helpful, harmless, and honest — and that design choice shows up in how it responds. Claude tends to be more cautious about strong claims, more willing to admit uncertainty, and noticeably better at handling very long documents in a single turn. For tasks like summarising a 100-page report, refactoring a large code file, or working through a nuanced ethical question, many professionals reach for Claude first.

This tutorial covers what Claude does differently, the XML-style prompting pattern Anthropic itself recommends, and the workflows that consistently get great results from the model.

2. How Claude Thinks Differently

Claude's training emphasises careful reasoning and conversational nuance. In practice you will notice three things:

• Long context, used well. Claude models have very large context windows — measured in hundreds of thousands of tokens — and they tend to use the full document rather than skimming the start and end. You can paste an entire book chapter and ask for a structured analysis.
• Strong instruction following. If you say "do not start your answer with a preamble", Claude usually obeys. If you provide a specific output schema, it tends to respect it down to the punctuation.
• Honest disagreement. Claude will push back on weak premises rather than confidently agree. That is a feature, not a bug — but it does mean you should frame requests carefully if you want a confident first draft.

Think of Claude as a careful senior consultant who reads everything you send, asks clarifying questions in their head, and then writes a thoughtful answer. The trade-off is that this carefulness can read as verbose if you don't tell Claude exactly how much to write.

3. The XML-Style Prompt Pattern

Anthropic's own documentation recommends using XML tags to clearly delimit sections of your prompt. Claude was trained to recognise these tags, and they noticeably improve response quality on complex tasks. The tags themselves are just plain text — you invent them as needed.

You are an experienced product editor for a B2B SaaS blog.

<task>
Rewrite the article below so it reads like a confident
explainer aimed at non-technical founders. Tighten by 20%.
</task>

<constraints>
- Keep all section headings.
- Do not invent statistics.
- Replace acronyms with plain English on first mention.
</constraints>

<article>
… paste the full article here …
</article>

<output_format>
Return only the rewritten article in Markdown. No commentary.
</output_format>

4. Using Claude's Long Context Well

Claude shines on document-heavy work. A few patterns to get the most out of the long context window:

• Paste the full document, not a summary. Claude can handle it, and you avoid losing detail in pre-processing.
• Ask for quoted evidence. "Quote the exact sentences that support each point" gives you something you can verify quickly.
• Index the document first. For very long material, ask Claude to first generate a table of contents with section labels. Then refer to those labels in follow-up turns.
• Use multiple documents. Wrap each in its own tag — <contract_a>…</contract_a>, <contract_b>…</contract_b> — and Claude will keep them separate when comparing.

5. Prompting Claude Projects and System Prompts

Claude offers Projects — persistent workspaces where you can attach files, set a custom system prompt, and chat with that context preloaded. The system prompt for a Project should follow the same XML-style discipline as a single-turn prompt.

You are the resident editor for a UK-based design studio.

<voice>
Friendly, direct, never salesy. UK English spelling
(colour, behaviour, analyse). Active voice.
</voice>

<defaults>
- Short paragraphs (max 3 sentences).
- Use H2/H3 Markdown headings.
- End every long piece with a "Next actions" bullet list.
</defaults>

<refuse>
- Do not write client testimonials.
- Do not invent case study numbers — ask the user instead.
</refuse>

6. Patterns That Work Especially Well on Claude

• "Think step by step in <scratchpad>." Ask Claude to reason inside a scratchpad tag, then write the final answer outside it. This dramatically improves reasoning on tricky questions and lets you read its working.
• The deliberate critique. "First, list three weaknesses of the draft I just gave you. Then propose fixes for each." Claude is comfortable being critical and will give substantive notes.
• Multi-document comparison. Provide two policies, contracts, or essays in separate tags and ask Claude to produce a clause-by-clause comparison table.
• Tone calibration up-front. Claude defaults to a slightly formal voice. State your desired tone explicitly — "punchy, conversational, no corporate language" — and it adapts cleanly.

7. Practice Exercises

8. Key Takeaways