Module 09 · Lesson 01
What Claude Actually Is (and Isn't)
Reading time: 16 minutes Track: Claude Fluency for Teams · Required for all learners
The wrong mental model costs you weeks
Most people arrive at Claude with one of two broken mental models.
The first is the search engine model: Claude is a smarter Google. You ask a question, it retrieves an answer. This leads people to write terse, keyword-heavy prompts and then complain that Claude "doesn't understand" them. It also leads to misplaced trust — if Claude is "looking things up," surely the facts are accurate?
The second is the magic oracle model: Claude knows everything and reasons perfectly. You just need to phrase the question right and the wisdom flows. This leads to over-reliance, skipped verification, and spectacular failures at exactly the moment it matters most.
Neither model is correct. And the actual model — once you have it — makes Claude dramatically more useful.
What Claude actually is
Claude is a large language model (LLM). Practically, here's what that means:
It predicts the most useful next token. Claude was trained on a vast corpus of human-generated text. At inference time, it generates responses by predicting, step by step, what text would be most useful given your prompt and its training. It is not retrieving stored facts. It is not reasoning through a formal logic engine. It is generating fluent, contextually appropriate text based on statistical patterns learned from training.
This has a critical implication: Claude does not "know" things in the way you know your own phone number. It has absorbed patterns from documents that contained accurate information, but it cannot cite a source, check its own outputs against a database, or flag when it's generating plausible-sounding text that happens to be wrong.
It has a knowledge cutoff. Claude's training data ends at a specific date. It has no access to the internet unless given a tool that provides it. Events, papers, pricing, or software versions after the cutoff are unknown to it — unless you provide that context in your prompt.
It has no persistent memory by default. Each conversation starts fresh. Claude does not remember what you discussed yesterday unless you (or your tooling) provides that context. This surprises people constantly.
It is stateless within a session — not across sessions. Within a single conversation, Claude does remember earlier messages. Across sessions, it does not.
What Claude is genuinely excellent at
Given the above, here is where Claude reliably delivers:
- Drafting and editing — generating first drafts, restructuring prose, improving clarity, matching tone
- Summarization — distilling long documents, meeting notes, research papers into key points
- Code generation and review — writing functions, explaining code, suggesting improvements, generating tests
- Structured analysis — breaking down a problem, comparing options, identifying tradeoffs
- Format transformation — converting unstructured notes to structured docs, tables to prose, code comments to documentation
- Brainstorming — generating options, considering angles, surfacing questions you hadn't thought to ask
Where to verify, every time
Claude can produce fluent, confident-sounding text that is factually wrong. This is not a bug to be fixed in a future version — it is a fundamental property of how language models work. The areas requiring systematic verification:
- Specific facts: dates, statistics, names, citations, legal citations, regulatory specifics
- Recent events: anything after the training cutoff
- Arithmetic and calculations: Claude often gets these wrong, especially multi-step
- Your own codebase: Claude doesn't know your internal APIs, naming conventions, or architecture unless you tell it
- Anything with legal or compliance consequences: treat Claude output as a draft requiring expert review
The correct mental model
Think of Claude as a brilliant, well-read colleague who has never worked at your company.
They have read extensively. They can write well, think through problems, generate options, draft documents, and explain concepts clearly. But they don't know your internal systems, they may have outdated information on fast-moving topics, they occasionally misremember details with great confidence, and they need context about your specific situation before their advice is useful.
When you treat Claude this way — providing context, checking important facts, using it as a force-multiplier on your own judgment rather than a replacement for it — the results are exceptional.
When you treat it as an oracle, you get burned.
One practical takeaway
Before you send any prompt to Claude, ask: "What would I need to tell a smart new hire to get a useful answer here?"
That's roughly the context Claude needs from you.