About this course. This course is independent of Anthropic. It is not produced, endorsed or accredited by Anthropic, and it reproduces none of Anthropic's course or documentation text. Claude's features change often, so check Anthropic's own help pages when something here looks different from what you see.
What Claude is
Claude is an AI assistant made by Anthropic. You type a request in plain language and it writes back: a summary, a draft email, an analysis, a list of ideas.
Under the surface, Claude is a large language model. It learned patterns from a very large amount of text, and it answers by generating the words most likely to follow from your request and everything else in the conversation.
Three consequences follow:
- It generates, it does not look up. Unless you give Claude a document or a tool, it writes from learned patterns, not from a database of verified facts.
- Its training has a cut-off date. It can be unaware of recent events or changes.
- It always sounds confident. Polished wording comes from the same process whether the content is right or wrong.
Treat Claude like a fast, well-read new colleague who has never seen your company's files and who rarely says "I'm not sure" unless asked.
Limit 1: Hallucination
A hallucination is an answer that is incorrect or invented but presented as fact. Anthropic's own help pages state that Claude can occasionally do this, and that users should not treat it as a single source of truth.
The highest-risk content is specific detail:
- Figures, dates and percentages
- Quotes and citations
- Names of people, laws, products and reports
- Anything recent
Example (finance): You ask for the current VAT threshold for a filing. Claude gives a precise figure in a tidy sentence. The figure may be last year's, or wrong. The quality of the sentence tells you nothing about the quality of the number.
What to do:
- Treat every specific fact as unverified until you check it against a source you trust.
- Ask Claude: "Which parts of this answer are you least sure of?"
- For high-stakes work (legal, financial, HR decisions), have a qualified person review before it is used.
Limit 2: The context window
The context window is the amount of text Claude can work with at one time. It includes your messages, Claude's replies and any files you attach. Think of it as a desk: everything on the desk is visible, and anything that does not fit is not.
Size is measured in tokens, which are chunks of text roughly the size of a short word. According to Anthropic's help centre, the baseline on paid plans is 200,000 tokens (about 500 pages of text), and newer models offer more, up to 1 million. These numbers depend on your plan and model, and they change, so check the current help page instead of memorising them.
Two points matter more than the number:
- A new conversation starts with an empty desk. By default, Claude does not carry over what you said in a different chat.
- Long conversations degrade. On paid plans with the code execution setting enabled, Claude may summarise earlier messages automatically to keep going. A summary keeps the gist and can lose detail.
Signs a conversation or document has outgrown the window:
- Claude ignores an instruction you gave early on, such as "UK spelling" or "under 200 words".
- It contradicts a decision you agreed earlier.
- It is accurate on the start and end of a long document but vague on the middle.
- You see a notice that the conversation has reached its length limit, or that earlier messages are being summarised.
Example (project management): After forty messages on a rollout plan, Claude schedules a task for a supplier you ruled out at the start. That is a context problem, not a reasoning problem.
What to do: restate the key facts and rules in your next message, or start a new conversation with a short summary.
Limit 3: Data sensitivity
Whatever you paste leaves your computer and is processed on Anthropic's systems. How it is handled depends on the plan:
- Personal plans (Free, Pro, Max): chats can be used to improve Claude if the model improvement option in your privacy settings is on. Check that setting.
- Work plans (Team, Enterprise): by default, inputs and outputs are not used to train models. Sending feedback such as a thumbs up or down is an exception.
Your employer's policy applies on top of either.
Before you paste, follow these steps:
- Identify what the content is and who it is about.
- Sort it into one of the three categories below.
- If it is not clearly safe, check your organisation's AI or data policy.
- Remove or replace names and identifiers you do not need.
- Paste only the minimum the task requires.
| Category | Meaning | Examples |
|---|---|---|
| Safe to paste | Public or harmless | A published job advert; your own meeting agenda with no names |
| Check policy first | Internal or confidential | An unreleased quarterly sales forecast; a client contact list |
| Do not paste | Secrets and regulated personal data | Passwords; employee medical absence records with names |
"Do not paste" holds unless your organisation has explicitly approved Claude for that type of data.
Do this now
- Open Claude and ask a factual question about a subject you know thoroughly, such as your own job, industry or home town. Request specific figures or names.
- Mark each specific claim as correct, wrong or unverifiable.
- Ask: "Which of those claims are you least confident about?" Compare the reply with your marks.
- List five items from your current work and sort each into the three categories.
Check your result: you should have at least one claim you had to verify outside Claude, and each of your five items should have one category and a one-line reason. If any item is in "check policy first", you should know which policy or person to ask.
Key takeaways
- Claude generates likely text. Fluency is not evidence of accuracy.
- Verify specifics: numbers, dates, names, quotes and anything recent.
- The context window is finite. Forgotten instructions and contradictions signal that you should restate or restart.
- Sort data before you paste: safe, check policy first, or do not paste.
- Plan type and privacy settings affect how your data is used, and your organisation's policy always applies.