Exercises

Practice designing an agentic loop with tool integration, structured error handling, and escalation patterns.

1. 1Define 3-4 MCP tools with detailed descriptions that clearly differentiate purpose, inputs, and boundaries. Include at least two similar tools requiring careful description to avoid selection confusion.
2. 2Implement an agentic loop that checks stop_reason to decide whether to continue tool execution or present the final response. Handle both "tool_use" and "end_turn" correctly.
3. 3Add structured error responses: errorCategory (transient/validation/permission), isRetryable boolean, and human-readable descriptions. Test that the agent retries transient errors and explains business errors.
4. 4Implement a programmatic hook intercepting tool calls to enforce a business rule (e.g., blocking operations above a threshold), redirecting to an escalation workflow.
5. 5Test with multi-concern messages and verify the agent decomposes the request, handles each concern, and synthesizes a unified response.

Practice configuring CLAUDE.md hierarchies, custom slash commands, path-specific rules, and MCP server integration.

1. 1Create a project-level CLAUDE.md with universal coding and testing standards. Verify project-level instructions apply across all team members.
2. 2Create .claude/rules/ files with YAML frontmatter glob patterns (e.g., paths: ["src/api/**/*"], paths: ["**/*.test.*"]). Test that rules load only when editing matching files.
3. 3Create a project-scoped skill in .claude/skills/ with context: fork and allowed-tools restrictions. Verify it runs in isolation without polluting the main conversation.
4. 4Configure an MCP server in .mcp.json with environment-variable expansion for credentials. Add a personal server in ~/.claude.json and verify both are available simultaneously.
5. 5Test plan mode versus direct execution on a single-file bug fix, a multi-file library migration, and a new feature with multiple valid approaches. Observe when plan mode provides value.

Practice designing JSON schemas, using tool_use for structured output, validation-retry loops, and batch processing.

1. 1Define an extraction tool with required and optional fields, an enum with an "other" + detail pattern, and nullable fields. Process documents where some fields are absent and verify the model returns null rather than fabricating values.
2. 2Implement a validation-retry loop: on validation failure, send a follow-up including the document, failed extraction, and specific error. Track which errors are resolvable (format) vs not (info absent).
3. 3Add few-shot examples demonstrating extraction from varied formats (inline citations vs bibliographies, narrative vs tables) and verify improved handling.
4. 4Submit a batch of 100 documents via the Message Batches API, handle failures by custom_id, resubmit with modifications (e.g., chunking oversized docs), and calculate processing time relative to SLA constraints.
5. 5Have the model output field-level confidence scores, route low-confidence extractions to human review, and analyze accuracy by document type and field.

Practice orchestrating subagents, managing context passing, error propagation, and synthesis with provenance tracking.

1. 1Build a coordinator delegating to at least two subagents (web search, document analysis). Ensure allowedTools includes "Task" and each subagent receives findings directly in its prompt.
2. 2Implement parallel subagent execution via multiple Task calls in a single response. Measure latency improvement over sequential.
3. 3Design structured subagent output separating content from metadata: each finding includes a claim, evidence excerpt, source URL/name, and publication date. Verify synthesis preserves attribution.
4. 4Simulate a subagent timeout and verify the coordinator receives structured error context, can proceed with partial results, and annotates coverage gaps.
5. 5Test with conflicting source data and verify synthesis preserves both values with attribution rather than arbitrarily selecting one, distinguishing well-established from contested findings.