Prompting Claude: the basics that actually matter

Claude responds to clear structure, explicit format and honest context. Three before-and-after examples show the difference, plus the two Claude-specific tricks worth knowing.

In this lesson you will

Build a prompt from the five parts that get first-time answers
Rewrite a vague request into a precise one (three worked examples)
Use XML-style tags and examples the way Anthropic's own guide recommends

Good prompting is cheaper prompting. A precise request costs a few hundred extra tokens and saves the two or three retries that cost thousands — or, on the free tier, the retries that push you into the limit. The structure below is what Anthropic's own prompting guide recommends, compressed to what beginners actually need.

Figure 1.Role, task, context, format, constraints. Most disappointing answers are missing the context or the format.

Example 1: writing

Before	After
Write an email to my landlord about the broken heater.	You are helping a tenant write a firm but polite email. Context: heater has been broken 9 days, two calls unanswered, lease clause 7 requires repair within 7 days (pasted below). Task: request repair within 48 hours and state I will contact the tenancy authority otherwise. Format: under 150 words, no legal threats beyond that one sentence. Constraint: plain language, no exclamation marks.

The 'after' version is longer to write and produces a usable email on the first try. The 'before' version produces a generic draft you then spend three turns fixing — three turns that each re-read the whole conversation.

Example 2: analysis

Before	After
Which AI plan should I get?	Here is my last month's usage: about 25 chats a day, mostly short, two long research sessions a week with PDFs, no coding. I am on Claude Free and hit the limit most afternoons. Compare Claude Pro against using the API through a third-party client. Output: a two-column table of monthly cost and friction, then a one-sentence recommendation. Assume the prices in this pasted table.

Example 3: coding

Before	After
Fix this code.	The function below should return the cheapest model for a given token count but returns nothing for counts above 100k. Language: TypeScript. Find the bug, explain it in two sentences, then return the corrected function only — no other changes, keep the existing style.

Two Claude-specific tricks

Tags for structure. Claude was trained to respect XML-style tags. Wrap pasted material in <document> … </document> and instructions in <instructions> … </instructions> and it will stop confusing the two — which matters when the document itself contains instructions.
Show, don't describe. One or two examples of the output you want (<example> … </example>) beat a paragraph describing it. This is the single biggest quality lever for formatting tasks like extraction and rewriting.

textA reusable prompt skeleton for Claude

You are a <role>.

<instructions>
<task in one sentence>
Output format: <table / bullets / JSON / word limit>
Constraints: <what to avoid>
</instructions>

<document>
<paste the material here>
</document>

<example>
<one example of the exact output shape you want>
</example>

Knowledge check

Claude keeps following instructions that were inside a document you pasted, not yours. What is the fix?

Lesson FAQ

Should I be polite to Claude?

It does not change the answer quality, and it costs a few tokens. Clear beats courteous. That said, a natural tone is fine; just do not pad.

How long can a prompt be?

Up to the context window — a million tokens on current Sonnet, Opus and Fable. In practice, put long material in a project or the system prompt so it is not re-sent every turn.

Does the same prompt work on ChatGPT and Gemini?

Mostly. The five-part structure is universal. The XML-tag habit is Claude-specific; other models tolerate it but were not trained on it the same way.