Capstone · Skills through practice

Hands-On Labs

Knowledge sticks when you build. These are the Exam Guide's four preparation exercises, expanded into concrete labs with self-check questions. The exam rewards judgment from experience — do at least one end-to-end.

Why bother when the exam is multiple-choice?The distractors are written by people who watched practitioners fail these exact tasks. Having felt a prompt-based rule fail, or a subagent lose context, makes the right answer obvious under time pressure. Fluency (recognizing an answer) is weaker than storage strength (having done it).

Lab 1 · Multi-Tool Agent with Escalation Logic

Reinforces D1 · D2 · D5

Define 3–4 MCP tools with detailed descriptions; include two deliberately similar tools that need careful wording to disambiguate.
Implement an agentic loop that branches on stop_reason — handle both "tool_use" and "end_turn" correctly.
Add structured errors (errorCategory, isRetryable, human-readable message); verify the agent retries transient errors and explains business errors.
Add a tool-call interception hook enforcing a business rule (e.g., block operations over a threshold) and redirect to escalation.
Send a multi-concern message; verify the agent decomposes, handles each concern, and synthesizes one response.

Success check: flip the hook off and re-run the over-threshold case — watch the prompt-only version fail intermittently. That failure is the lesson behind "use a guarantee."

Lab 2 · Claude Code for a Team Workflow

Reinforces D3 · D2

Create a project-level CLAUDE.md with coding + testing standards; confirm it applies for a second "teammate" checkout (not user-level).
Create .claude/rules/ files with paths globs (e.g., ["src/api/**/*"], ["**/*.test.*"]); confirm they load only when editing matching files.
Create a project skill with context: fork and allowed-tools; confirm it runs isolated without polluting the main conversation.
Configure an MCP server in .mcp.json with ${ENV} expansion; add a personal one in ~/.claude.json; confirm both are available at once.
Try plan mode vs direct on three tasks: a one-file bug fix, a multi-file migration, a feature with multiple valid designs. Note where plan mode earns its cost.

Success check: put an instruction in ~/.claude/CLAUDE.md and confirm a fresh clone doesn't see it — the exact failure behind the "new teammate" question.

Lab 3 · Structured Data Extraction Pipeline

Reinforces D4 · D5

Define an extraction tool whose JSON schema has required + optional + nullable fields and an enum with "other"+detail. Feed a doc missing a field; confirm it returns null, not a fabrication.
Build a validation-retry loop (Pydantic/JSON schema): on failure resend the doc + failed extraction + specific error. Track which failures retry fixes (format) vs can't (absent info).
Add few-shot examples for varied structures (inline citations vs bibliographies; tables vs narrative); confirm improved consistency.
Submit a 100-doc batch via the Message Batches API; handle failures by custom_id (chunk oversized docs); compute processing time vs an SLA.
Output field-level confidence; route low-confidence to human review; analyze accuracy by document type and field.

Success check: make a schema field required and watch fabrication return — then make it nullable and watch it stop. That's the anti-hallucination rule in your hands.

Lab 4 · Multi-Agent Research Pipeline (design & debug)

Reinforces D1 · D2 · D5

Build a coordinator delegating to ≥2 subagents; ensure allowedTools includes "Task" and each subagent gets findings in its prompt (no automatic inheritance).
Run subagents in parallel via multiple Task calls in one response; measure the latency win vs sequential.
Give subagents structured output separating content from metadata (claim, excerpt, source, date); confirm synthesis preserves attribution.
Simulate a subagent timeout; confirm the coordinator receives structured error context and can proceed with partial results + coverage-gap annotations.
Feed conflicting source stats; confirm synthesis keeps both with attribution (and dates) rather than picking one.

Success check: deliberately over-narrow the coordinator's decomposition and watch coverage gaps appear even though every subagent "succeeds" — the root-cause question, reproduced.

Quick self-check

In Lab 1 you disable the interception hook and rely only on a system-prompt rule to block over-threshold operations. The observed behavior best illustrates that:

Prompt-based rules have a non-zero failure rate; guarantees need deterministic enforcement
System prompts are ignored entirely whenever any tool is available
Interception hooks are only necessary when the model lacks tools
Business rules should always be enforced by lowering the temperature

Option 1. The whole point of the lab: probabilistic prompt compliance fails some fraction of the time; deterministic hooks/gates don't. This is the exam's master heuristic.

Ask your teacher. Want starter scaffolding for any lab (tool schemas, a loop skeleton, a rules-file template)? Ask and I'll generate it. No cloud access needed — you can run these locally with the Agent SDK / Claude Code.