Claude Architect Foundations Course Notes

Claude architecture starts with choosing the right surface for the job. The direct API gives an application full control over prompts, message history, tools, retries, storage, and user experience, while Managed Agents are a better fit when Anthropic-managed sessions and built-in autonomy reduce the amount of orchestration the product team must build.

• Treating the newest Claude model as the automatic answer instead of matching model capability, latency, and cost to the workload.

• If the scenario emphasizes application-owned conversation state, custom orchestration, or exact API behavior, think Messages API before Managed Agents.

Claude API design begins with authentication, workspace separation, and request sizing. Production systems should keep API keys out of client-side code, separate environments or use cases with workspaces where appropriate, and handle large prompts, files, and rate behavior before the first user-facing rollout.

• Putting durable behavior in a user message instead of the Messages API `system` field.

• If the scenario asks how to preserve conversation context, the application must send the relevant message history on each request.

Prompting architecture separates durable behavior from one-off user input. Put role, tone, safety constraints, and output contract in the system prompt, then keep the user turn focused on the specific task and source material for that request.

• Using a longer prompt when the real issue is missing success criteria, weak output format, or unclear source boundaries.

• If the prompt has policy text, examples, and user-provided material, XML tags are usually the cleanest way to separate them.

Tool architecture starts with a precise contract. Each tool needs a name, description, and input schema that tells Claude when the tool is appropriate and tells the application exactly what arguments it may receive.

• Thinking Claude executes client tools directly; the application must run the tool and return a `tool_result`.

• If the question asks what a tool definition needs, look for name, description, and input schema rather than application code.

Model Context Protocol is an integration standard for connecting AI clients to external tools and data sources. In Claude Code, MCP is not a separate model or a mission-control layer; it is the protocol used to expose approved capabilities from MCP servers to an MCP client.

• Expanding MCP as a made-up platform name instead of Model Context Protocol.

• If the scenario connects Claude Code to external tools, repositories, databases, or services, map it to MCP clients and MCP servers.

Claude Code is most useful when the task needs repository context plus action. It can inspect files, explain unfamiliar modules, edit code, run commands, and use development tools, so the architect's decision is how much execution authority to give it for the current repository and risk level.

• Using normal edit mode for broad risky changes when Plan Mode should inspect and propose before modifying files.

• If the task is exploratory or high-risk, Plan Mode is the safer Claude Code workflow before execution.

Least privilege should shape every Claude Code deployment. Give users and agents only the repository, command, tool, and external-system access needed for the task, then expand permissions deliberately when the risk is understood.

• Treating approval prompts as inconvenience instead of the boundary between Claude intent and real system changes.

• If a command can change files, call networks, install packages, or touch credentials, expect an approval gate or managed permission policy.

Claude Projects organize work around a durable context boundary. A project can hold instructions, knowledge, and shared materials for a product, client, team, or initiative so users do not have to rebuild the same context in every conversation.

• Using one project for unrelated customers or teams and accidentally mixing instructions, files, and conversations.

• If the scenario needs persistent shared context for a team or initiative, Claude Projects are the likely answer.