Lesson 1: Course Ground Rules and How Claude Reads Text

Ground rules for this course

This course is independent of Anthropic. Anthropic did not write, review or endorse it, and it reproduces none of Anthropic's course or documentation text. Treat it as a practitioner's guide and treat Anthropic's own pages as the authority on what Claude can do today.

That distinction matters because Claude changes fast. Models, features, limits and plans are added, renamed or retired within weeks. Anything you read here, in a blog post or in a video may already be out of date.

How to verify a Claude claim

Do this whenever a feature name, limit, price or plan matters to your work.

1. Write down the exact claim. Example: "Model X supports feature Y on my plan."
2. Open the matching Anthropic page:
   • Models, IDs and limits: the models overview in the Claude Platform docs (platform.claude.com/docs).
   • API changes: the Claude Platform release notes.
   • Chat app changes: the Claude apps release notes in the Claude Help Center (support.claude.com).
   • Retired models: the model deprecations page.
3. Read the status label. Anthropic's docs mark features as generally available, beta or research preview. Beta features can change or be removed, and on the API they often need a special beta header.
4. Check the date of the page or release note entry.
5. Record what you found. If you cannot confirm a claim on an Anthropic page, label it "unconfirmed" and do not build on it.

This course follows the same rule: it will say when something is beta or unconfirmed.

What a token is

Claude does not read letters or words. Before the model sees your text, software called a tokenizer cuts it into tokens: small units that can be a whole word, part of a word, a single character or punctuation.

Think of Lego bricks. Common words are a single large brick. Rare words, names, code and typos are built from several smaller bricks, which lets the model handle text it has never seen before.

An illustrative split (real splits vary by model):

• cat → 1 token
• tokenization → perhaps token + ization
• Zxqvly → several small pieces

Why token counts differ from word counts

Tokens and words do not map one to one. The count depends on:

• Word familiarity. Common words cost fewer tokens than rare ones.
• Language. The characters-per-token ratio changes between languages.
• Content type. Code, numbers, URLs and punctuation split differently from prose.
• The model. Different Claude generations use different tokenizers.

The last point is easy to miss. Anthropic's docs currently state:

• On models before Claude Opus 4.7, one token is roughly 3.5 English characters.
• Claude Opus 4.7 and later models use a newer tokenizer that produces about 30 percent more tokens for the same text. The exact increase depends on the content.
• On the current tokenizer, 1 million tokens is roughly 555,000 English words. On the earlier one it was about 750,000.

In practice that means roughly 1.8 tokens per English word on current models, against about 1.3 on older ones. These are estimates, and they will shift again when tokenizers change.

What to do:

• In chat: use the rule of thumb for a rough size check. A 1,000-word document is well over 1,000 tokens.
• On the API: use Anthropic's token counting endpoint. It is free to use, has its own rate limits, and counts under the tokenizer of the model you name. Anthropic describes the result as an estimate that can differ slightly from actual usage.
• When you switch models: recount. Do not reuse a number measured on an older model.

Tokens matter because limits and API billing are measured in tokens, not words.

How Claude writes a reply

Claude is built on a language model trained to predict what comes next in text. A reply is produced in this order:

1. Everything supplied for this turn is assembled into one sequence of text and converted to tokens.
2. The model processes the whole sequence.
3. It produces one next token.
4. That token is added to the end of the sequence.
5. Steps 2 to 4 repeat until the reply is complete.

Two consequences follow:

• Claude works only from the text it is given. It cannot see what is in your head, on your screen or in a file you did not supply. If a detail matters, it must be in the text.
• The reply is built front to back. Each token is chosen in light of everything before it, including the reply so far. Claude does not draft a whole answer and then paste it in.

A useful comparison is phone autocomplete with far greater capability: it also extends text one piece at a time. The difference is that Claude considers the entire text each time, not just the last few words.

Try it now: take a prompt you used recently. Ask, "If I could see only this text, would I have everything needed to do the task?" List what is missing.

Key takeaways

• This course is independent of Anthropic. Anthropic's current pages are the authority.
• Before relying on any feature, name, limit or plan, check it on Anthropic's docs or release notes and note whether it is generally available, beta, research preview or unconfirmed.
• A token is a chunk of text: a word, part of a word, a character or punctuation.
• Token counts differ from word counts and vary by language, content and model. Newer Claude models produce about 30 percent more tokens for the same text than models before Claude Opus 4.7.
• Claude generates a reply one token at a time, each time working from the full text it has been given and nothing else.