Claude for Lawyers: Tokens and Context Windows Explained

If you are a lawyer using Claude, ChatGPT, Claude Cowork, or Claude Code for legal work, you need to understand tokens.

Tokens are how large language models measure text. They affect usage limits, speed, cost, and the quality of the answers you get back. More importantly, tokens explain why a chat that worked well at the beginning can start getting worse after you add a long brief, a discovery demand, a few PDFs, and twenty follow-up questions.

I made a video walkthrough here:

You can also watch it on YouTube: Claude for Lawyers: Tokens and Context Windows Explained.

What Tokens Are

Claude and ChatGPT do not really read words the way people do. Under the hood, they convert text into tokens, and those tokens are what the model reads and writes.

A token is usually smaller than a word, but larger than a character. A simple rule of thumb is that one word is often around 1.25 tokens, but the exact number depends on the text.

You can see this visually with OpenAI's tokenizer: platform.openai.com/tokenizer.

For normal users, the practical rule is simple: more text means more tokens.

Why Tokens Matter for Lawyers

Tokens matter for three reasons.

First, they affect usage limits. If you are using Claude Pro, Claude Max, ChatGPT, or another consumer AI product, your plan has some kind of usage cap. Long chats and large documents burn through that allowance faster.

Second, they affect speed and cost. If a model has to reread a long chat every time you send a new message, each response takes more work.

Third, and most important for legal work, they affect accuracy. As the chat gets longer, the model has more context to juggle. Eventually it starts forgetting details, missing instructions, and hallucinating more often.

Context Windows Are Not Infinite

Every model has a context window. That is the maximum amount of information the model can consider at one time.

Some models advertise very large context windows. But lawyers should not treat those limits as infinite.

There is a hard limit where the model literally cannot take more tokens. But there is also a softer practical limit where the model still responds, but its performance gets worse. In legal work, that softer limit matters more.

If you have ever noticed Claude or ChatGPT getting worse in the same conversation, this is often why. The model is not necessarily broken. The chat may just be too full.

Why Word Documents Can Be Token-Heavy

Word documents are not just text. A .docx file is really a package of XML files containing the document text plus formatting, styles, numbering, metadata, and other structure.

That matters because Claude or ChatGPT may need to process far more than the words you see on the page.

In the video, I compare the same short discovery demand across several formats:

• DOCX
• HTML
• Markdown
• plain text

The plain text and Markdown versions are far smaller. The DOCX version uses many more tokens because it contains a lot of formatting information the model usually does not need if the task is just to understand the substance.

This does not mean lawyers should stop using Word. Word is still the final drafting and filing artifact for most legal work. But if you are asking Claude to read, summarize, compare, or reason about a document, a cleaner format can often work better.

Practical Habits

For legal AI workflows, the basic habits are:

• Start new chats more often than feels necessary.
• Keep each chat focused on one task or one document set.
• Use plain text, Markdown, or HTML when formatting is not important.
• Use Word when the output needs to be a Word document.
• Avoid dumping unnecessary documents into the context window.
• Treat long chats as a source of accuracy problems, not just clutter.

The goal is not to obsess over token counts. The goal is to avoid wasting context on information the model does not need.

Where DocketDrafter Fits

Claude is good at reasoning and drafting, but serious legal drafting also needs controlled document structure.

A litigation document is not just loose text. It needs the right caption, signature block, formatting rules, prior work, reusable language, citation review, version history, and clean Word output.

DocketDrafter is the controlled drafting workspace around Claude: firm-owned playbooks, approved templates, reusable prior work, citation review, version history, and court-ready Word rendering.

In other words: Claude drafts. DocketDrafter controls the document.

If your firm is starting to use Claude seriously and wants to turn that workflow into a repeatable firm system, you can learn more at DocketDrafter.

You can also email me directly at tommy@docketdrafter.com.