Chapter 4 Claude Fundamentals and Product Context

4.1 Claude is a model family, not one fixed interface

Claude is a family of language models developed by Anthropic. The capabilities available in a session depend on both the selected model and the product that hosts it. Claude.ai, Claude Code, an API integration, and a third-party product can expose different tools, permissions, context, and data-retention policies.

Claude can work with text, code, and images. When a host provides tools, it may also search the web, read and edit files, execute commands or code, and connect to external systems. Do not infer available capabilities from the model name alone; inspect the current product, tool list, and permission settings.

Primary references:

4.2 A practical mental model

For most tasks, treat the system as four layers:

  1. Model — generates and interprets language, code, and structured content.
  2. Context — the instructions, conversation, files, and tool results visible during the current request.
  3. Tools — capabilities supplied by the host, such as file editing, Bash, web search, or code execution.
  4. Permissions and environment — boundaries controlling what tools can read, change, or communicate externally.

This distinction prevents common mistakes. For example, a chat session without a code tool cannot execute a program, while Claude Code can run project commands when its environment and permissions allow it.

4.3 Useful strengths

Claude is often useful for:

  • explaining unfamiliar code and technical concepts;
  • drafting, refactoring, and documenting code;
  • proposing tests, edge cases, and alternative designs;
  • synthesizing information supplied in the current context;
  • exploring a repository and carrying out multi-step work when tools are available; and
  • translating ambiguous goals into a concrete plan for human review.

These are capabilities, not guarantees. Output quality depends on the task, context, model, tools, and verification process.

4.4 Important limitations

4.4.1 Outputs can be wrong

Claude can produce plausible but incorrect facts, citations, assumptions, code, or calculations. Verify consequential claims against primary sources and verify code with tests, builds, static analysis, or direct inspection.

4.4.2 Product capabilities vary

Web access, code execution, file access, memory, and integrations are not universal. A feature may be unavailable, disabled, permission-gated, or supplied by a third-party host. State the environment when documenting a workflow.

4.4.3 Context is finite

Long conversations, large files, and verbose tool output consume context. As context fills, earlier detail may be summarized or receive less attention. Provide focused evidence, keep durable project rules in repository instructions, and begin a fresh session when the task changes substantially.

4.4.4 Current information requires retrieval

Model training knowledge is not a substitute for current documentation. For changing APIs, models, prices, regulations, or library behavior, use an enabled retrieval tool and cite an authoritative source. If retrieval is unavailable, label the information as unverified rather than guessing.

4.4.5 Tool use creates real risk

An agent with file, shell, browser, or external-service access can change state. Use least privilege, protect secrets, review destructive commands, preserve unrelated work, and require explicit approval for publishing or external actions.

4.5 A verification-centered partnership

Divide responsibilities deliberately:

Human responsibility Agent contribution
Define goals, risk tolerance, and factual constraints Explore, draft, compare, and implement
Supply domain and organizational context Surface assumptions and missing information
Approve consequential or external actions Run authorized checks and report evidence
Judge whether the result is useful and correct Iterate in response to test results and review

The agent should show evidence rather than merely announce success. Appropriate evidence might include a passing test, a clean build, a rendered screenshot, a query result, or a primary-source link.

4.6 Exercise: compare environments

Choose one task and try it in two Claude environments, such as a chat interface and Claude Code.

  1. Record which model and tools are visible.
  2. Ask each environment to explain a small local function.
  3. Ask whether it can run the function and verify the answer.
  4. Compare what each environment can actually inspect or execute.
  5. Write down which claims were generated and which were verified.

4.7 Key takeaways

  • Separate the model from the host product, tools, and permissions.
  • Never assume web, file, memory, or execution capabilities are present.
  • Treat fluent output as a candidate result until evidence validates it.
  • Use current primary documentation for changing product behavior.
  • Keep humans responsible for goals, risk, and consequential decisions.