Better Claude prompts come from structure and context, not clever phrasing. Modern Claude models take instructions literally and do exactly what you ask — so the fix is being explicit, organizing the parts of your prompt clearly, and telling Claude the reasoning behind the request. If you’re copying your ChatGPT habits over, a few of them are working against you.
Anthropic has been open about how their newer models behave. Since Claude Sonnet 4.5 through Sonnet 5, the models are instruction-first and literal: they don’t pad, they don’t guess at extra intent, they do the thing. That makes clear structure pay off more than it used to.
1. Use XML tags once the prompt gets long
For any prompt longer than about three sentences, wrapping each part in an XML tag is the highest-leverage move. When your prompt mixes instructions, background, examples, and the actual input, tags remove the ambiguity of where one ends and the next begins.
It looks like this: put your background inside <context>…</context>, the job inside <task>…</task>, and the thing to act on inside <input>…</input>. Claude parses tagged prompts far more consistently than a wall of plain text. You don’t need to tag a one-line question — this is for the meaty prompts where things get muddled.
2. Put role and context before the task
Order matters. Lead with who Claude should be and the background it needs, then give the task, then the output format. Context first lets Claude pick the right framework before it reads what you want — the same way you’d brief a person before handing them the job, not after.
A workable skeleton: role and context first, then the task, then specific instructions, then the output format. You don’t need every slot every time, but that sequence beats dropping the task in cold.
3. Explain why, not just what
Claude leans careful, and it responds well when you explain your reasoning instead of just issuing commands. "Rewrite this to be shorter" is fine. "Rewrite this to be shorter because it’s going in a mobile push notification with a 120-character limit" is better — now Claude optimizes for the real constraint, not a generic idea of "short".
4. Show an example when format matters
A few-shot example — one or two samples of the output you want — is one of the techniques that still clearly helps modern Claude. When you need a specific structure or voice, showing beats describing. Paste a good example and tell Claude to match it.
| If you’re used to… | For Claude, do this |
|---|---|
| Barking a blunt one-liner | Add the why — Claude uses the reasoning |
| A long unstructured paragraph | Wrap the parts in <context>, <task>, <input> tags |
| Dropping the request in first | Context and role first, then the task |
| Assuming it’ll infer your intent | Be literal — it does exactly what you write |
The one-line summary
Claude wants to be told clearly, in order, with the reasoning attached. That’s most of the gap between an okay Claude answer and a great one. And whether you’re on Claude, ChatGPT, or Gemini, the underlying discipline is the same — say the outcome, the context, and the format before you send. BeforePrompt flags the missing one right in the Claude prompt box, so you catch it before the reply, not after.