Best Claude Prompts

How Claude reads a long prompt

Most of the prompts on this page hand Claude something big: a contract, a transcript, a manuscript, a pile of support tickets. Anthropic's prompting guide is specific about how to lay out a request like that, and the blocks here follow it. You can do the same in anything you write yourself.

1. Material first, question last. For long inputs, Anthropic recommends putting the documents at the top and your question, instructions and examples underneath. A request that starts with «Find the risky clauses» and then pastes forty pages works, but it works better the other way round.
2. One tag per kind of thing. Wrap each input in its own tag with a plain name, like <contract>, <notes> or <draft>, and keep the names the same from one prompt to the next. With several documents, give each its own <document> tag and say where it came from inside it.
3. Quotes before conclusions. Ask Claude to pull out the passages that bear on your question first, then answer from those. It's Anthropic's own advice for long documents, and it gives you something to check against the source.
4. The reason behind a rule. «Keep it under 200 words, because it goes in a board pack that's already too long» gets followed more sensibly than the word count alone. Claude generalizes from the reason to cases you didn't list.
5. What you want, said positively. «Write in flowing paragraphs» steers better than «don't use bullet points». The guide makes the same point about formatting in general.

Here's that layout short enough to type from memory:

<document>
[paste the text]
</document>

I'm [who you are], and I'll use this to [decide what].
First, quote the passages that bear on [your question] in <quotes> tags.
Then answer from those quotes only, as [the shape you want].
If the document doesn't say, write "not stated".

Every block below is that layout worked out for one job. The Long Document Summarizer ties each key point to the section it came from, and the Research Paper Interrogator marks its own inferences so you can tell them apart from what the authors wrote.

Which prompt for which job

The third column is the tag each block expects. Paste your material between the opening and closing tag and leave the rest as it is.

Coding work in the terminal has its own collection, Claude Code prompts. If you keep pasting the same instructions, a skill may suit you better than a prompt; our Claude skills directory lists ready-made ones.

One request, before and after

We wrote this pair to show what layout does on a two-document job. Here's the version most of us type first, with both leases pasted underneath:

What changed in the new lease? [old lease] [new lease]

Claude has to work out where one lease ends and the other begins, and «what changed» covers every comma. You'll get a long list where the rent increase sits next to a renumbered heading. The second version takes a minute longer:

<old_lease>
[paste the current lease]
</old_lease>
<new_lease>
[paste the renewal]
</new_lease>

I'm the tenant and I have to sign within two weeks.
Quote every place where rent, deposit, notice period or repairs
wording differs, old version first. After each pair, one line:
better or worse for me, and why. Ignore formatting changes.

Now there's no doubt which text is which, the question comes after both documents, and each verdict sits under the quotes it rests on, so you can check it in seconds. The Document Version Diff does this for any pair of documents and counts the cosmetic changes without listing them.

Follow-up lines that fix an answer

One line in the same chat is usually enough.

If an artifact throws an error, there's a «Try fixing with Claude» button next to the message. It copies the error into a new message for you.

Where Claude answers go wrong

The prompts here are built against these, but it pays to know the signs.

• Charts in long PDFs go unread. Claude reads images, charts and graphics in PDFs of 100 pages or fewer. From 101 pages on it reads the text only, so a figure on page 140 doesn't exist for it. Upload the pages with the charts as a separate, shorter PDF.
• Word files lose their pictures. For documents other than PDFs, Claude extracts text only. A chart pasted into a DOCX is invisible.
• Chats in a Project don't share memory of each other. Anthropic's help center says context isn't shared across chats in a project unless it's in the project knowledge. The project's name and description aren't visible to Claude either, so put anything it needs in the instructions.
• Long chats get compressed. On paid plans with code execution on, Claude summarizes earlier messages as a chat nears its limit. Detail from the first hour can blur, so keep the source documents in the project instead of scrolled far back in a thread.
• Quotes drift. Even when asked for verbatim wording, check a quote against the document before you forward it. Searching for four or five words from it takes seconds.

Claude features that change the result

These are as of October 2026, from Anthropic's help center. They change often.

• How much it reads. On paid plans, Anthropic lists a 1M-token context window in chat for Claude Fable 5.1, Opus 5.5 and Sonnet 5.5, among others, and 200K for models outside that list.
• Uploads. Up to 20 files per chat, 500 MB each, and PDFs up to 1,000 pages. PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB and JSON all work; XLSX needs code execution and file creation switched on.
• Projects. Available on every plan; free accounts can have five. A project holds a knowledge base and its own instructions, which apply to every chat inside it, and that's where the Project Knowledge Curator output goes. Files you add to a project are capped at 30 MB each. On paid plans, Claude switches to retrieval when the knowledge nears the context window limit.
• Artifacts. Every plan can make them, but «Cloud code execution and file creation» must be on under Settings, then Capabilities. Paid plans also get templates for Docs, Slides and Design, in beta. Artifacts made in chat before September 16, 2026 are now legacy: they keep working and can still be shared, but new ones are made the new way.
• Setting how Claude writes. The help center lists three layers: instructions for Claude, which apply to every chat; project instructions; and skills, which it recommends for tone and format. If an older guide sends you looking for Styles, you won't find them on that page now; skills are what it points to for that job. Our explainer on how Claude skills work walks through making one.

Our Claude review keeps the plans and prices, and Claude vs ChatGPT covers when to open which.