30 CCAR‑F build exercises

Reading gets you recognition. Building gets you recall. Thirty things to actually ship — weighted to the exam blueprint, each with build steps, acceptance criteria and the trap it inoculates you against. Roughly 22 hours of work, free.

Open them in the study app All 54 lessons

Domain 1: Agentic Architecture & Orchestration

27% of exam · 8 exercises · ~5.9 h

Build a bare agentic loop from stop_reasonDrive a tool-using agent by hand so the loop stops being magic.Starter · ~40 min Handle pause_turn and resume a server-side toolResume a long-running server tool correctly instead of restarting it.Core · ~35 min Return parallel tool results without breaking pairingFan out several tools in one turn and pair every result correctly.Core · ~45 min Implement chaining, routing and parallelisation side by sideFeel why the three basic workflow patterns are not interchangeable.Core · ~50 min Build an orchestrator-workers system with an evaluator loopDecompose at runtime, then iterate against an explicit rubric.Stretch · ~60 min Scope a subagent's context and tools deliberatelyProve that subagents start clean, and design the handoff accordingly.Core · ~45 min Checkpoint an agent so a crash does not restart itPersist enough external state to resume mid-run.Core · ~40 min Classify errors and decide what propagatesStop a subagent failure from silently corrupting a coordinator's answer.Core · ~40 min

Domain 2: Claude Code Configuration & Workflows

20% of exam · 6 exercises · ~4.2 h

Map settings precedence by making rules collideLearn the precedence order by observing it, not by memorising a list.Starter · ~35 min Enforce a rule with a hook, not with CLAUDE.mdSee the difference between guidance and enforcement, concretely.Core · ~45 min Design a CLAUDE.md hierarchy for a monorepoPut each instruction at the scope where it belongs.Core · ~40 min Load rules conditionally with path-scoped frontmatterKeep specialised instructions out of context until they are relevant.Core · ~35 min Package a repeatable workflow as a SkillTurn a procedure you repeat into something invocable and versioned.Core · ~45 min Run Claude Code headless in CI with scoped permissionsAutomate a review job that cannot escape its sandbox.Stretch · ~55 min

Domain 3: Prompt Engineering & Structured Output

20% of exam · 6 exercises · ~4.3 h

Get guaranteed JSON two different waysSeparate structured outputs from tool-use schema enforcement.Starter · ~35 min Break a schema that validates but is wrongInternalise that schemas prevent syntax errors, not semantic ones.Core · ~45 min Build a production prompt with XML structure and few-shot examplesMake a prompt legible to the model and maintainable by you.Core · ~40 min Process a backlog with the Message Batches APIUse batch where it belongs and know where it does not.Core · ~45 min Replace prefill and legacy sampling parametersUpdate habits that current models reject or ignore.Core · ~40 min Build a prompt evaluation harness with regression detectionMake prompt changes safe to ship.Stretch · ~55 min

Domain 4: Tool Design & MCP Integration

18% of exam · 5 exercises · ~3.9 h

Rewrite a bad tool schema into a good oneLearn that tool descriptions are the primary lever on selection quality.Starter · ~40 min Return errors an agent can actually act onDesign tool failure output as carefully as tool success output.Core · ~45 min Build an MCP server exposing all three primitivesKnow tools, resources and prompts by having chosen between them.Core · ~55 min Make tool calls idempotent and safe to retrySurvive a partial success without duplicating side effects.Core · ~45 min Scale past a bloated tool surfaceFix the failure mode where too many tools degrade selection.Stretch · ~50 min

Domain 5: Context Management & Reliability

15% of exam · 5 exercises · ~3.8 h

Place cache breakpoints and verify the hitCut cost with prompt caching and prove it worked.Core · ~45 min Compare context editing, compaction and external memoryPick the right strategy for a conversation that outgrows the window.Core · ~45 min Measure long-context degradation on your own dataSee position effects rather than take them on faith.Core · ~40 min Build a RAG pipeline with citations and grounding checksMake every claim traceable to a retrieved source.Stretch · ~60 min Define escalation thresholds and calibrate confidenceDecide in advance when the system must stop and ask a human.Core · ~40 min