Cram Sheet — The Day-Before One-Pager

Read every question with the five-question actor test → who decides? model proposes, harness executes · who runs it? model-CHOSEN, harness-RUN · who guarantees it? persuaded / API-enforced / harness-validated · whose process executes? dispatcher / MCP server / Anthropic infra · who manages memory? physics (window, attention) = model; contents = harness. When two options “work”, pick the one with every verb on its rightful actor.

Deep breath — you know more than you think. This page is deliberately terse: every line is a compressed spring that should unfold into a full explanation in your head. If one doesn't, that's your last-day study target (the dotted terms open cards).

The one sentence

The model decides; the harness executes — every correct answer moves enforcement into deterministic layers (gates, hooks, schemas, allowlists, budgets) and keeps prompts for steering.

Blueprint (CCAR-F)

Agentic 27 · Claude Code 20 · Prompting/Structured 20 · Tools/MCP 18 · Context/Reliability 15 — 60 Q · 4 of 6 scenarios · 120 min · 720/1000 (= passing scaled score).

Tables to hold in your head

Stop reasons → end_turn deliver · tool_use execute+loop · max_tokens TRUNCATED-never-parse · stop_sequence custom · refusal → route to a human.

Workflow patterns → chain (fixed sequence) · route (classify→dispatch) · parallel-section (independent) · parallel-vote (confidence) · orchestrator–workers (dynamic decomposition) · evaluator–optimizer (criteria + iteration). Workflows before agents when the path is known.

MCP primitives → tools = model-controlled · resources = app-controlled · prompts = user-controlled. Client:server = 1:1 stateful. stdio local · Streamable HTTP remote. Value = N×M → N+M.

Claude Code surface → CLAUDE.md memory (project in git > user personal; specific wins) · skills = model-CHOSEN (by description), harness-RUN expertise · slash commands = user-invoked templates ($ARGUMENTS) · subagents = isolated context + tool policy · hooks = deterministic (PreToolUse exit-2 blocks; PostToolUse transforms) · settings precedence enterprise > CLI > local-project > shared-project > user · headless claude -p + stream-json for CI.

Structured output ladder → prompted JSON (advisory) < prefill { < schema-as-forced-tool (tool_choice, works on every model) < native structured outputs / strict tool inputs (the strongest shape guarantee, where supported) < harness validation (closed schema) + one repair retry. Prefill is incompatible with extended thinking; prefill is model-dependent on newer models.

Injection defense → bound input → quarantine as data → deterministic output gates. Layer 3 is the one that holds.

Caching → byte-identical prefix; stable content BEFORE the breakpoint, volatile after; reads ≈10% price; loops amortize.

Reliability → 429 (rate limit) / 529 (overloaded) = backoff+jitter · idempotency keys for side effects · checkpoints resume, never replay · degradation ladder ends in honest failure · Batch API for offline volume · pin model versions in prod · instrument before treating.

Multi-agent → subagents exist for CONTEXT ISOLATION; return summaries; parallelize reads, serialize writes; hub-and-spoke; know when one loop suffices.

The six scenarios in one line each

S1 support: escalation is a designed terminal path; money = gate+HITL. S2 team Claude Code: guarantees = hooks, standards = project CLAUDE.md. S3 research: coordinator + summary-returning subagents. S4 dev tooling: scoped Bash allowlists; Grep/Glob read-only. S5 CI/CD: headless + least privilege; block on determinism only. S6 extraction: forced schema-tool + stop_reason + closed schema.

Exam mechanics

~90 s/question after 1 min/scenario read · flag-and-move · "most appropriate" = Anthropic's published guidance (simplest thing, workflows-first, deterministic enforcement) · re-read the constraint sentence before answering · 10-min review pass.

Night-before checklist

Pearson VUE system test done · ID ready · wired internet · clean desk · water · sleep > cramming past this page.

Named features added from the official exam guide (v0.2) — last-pass list

• Task tool spawns subagents; coordinator's allowedTools must include "Task"; subagents inherit NO context (pass findings in the prompt); parallel = multiple Task calls in ONE response.
• --resume <name> = continue named session (tell it what files changed); fork_session = parallel branches off a shared baseline; stale tool results → fresh session + injected summary.
• Plan mode = architectural/multi-file/multiple-valid-approaches; direct execution = scoped single-file fixes; Explore subagent = quarantine verbose discovery.
• .claude/rules/ + YAML paths: globs = conventions that follow file patterns across the tree (test files everywhere); nested CLAUDE.md = one subtree only; skill frontmatter: context: fork (isolated context), allowed-tools (restrict), argument-hint (prompt for args).
• CI flags: claude -p (non-interactive), --output-format json, --json-schema (structured findings); review with an INDEPENDENT session, feed prior findings to dedup.
• Batches API: 50% cost · ≤24h · no SLA · NO multi-turn tools · custom_id correlation · resubmit failures only · submit every ~4h to hit a 30h SLA · never for blocking pre-merge checks.
• MCP errors: isError + errorCategory (transient/validation/business/permission) + isRetryable + human description; access failure ≠ valid empty result; local retry first, propagate with partial results.
• Schemas: nullable fields prevent fabrication; enum + "other"+detail + "unclear"; tool_use kills syntax errors, harness catches semantic ones (calculated_total vs stated_total).
• Context: lost-in-the-middle → key findings at TOP + section headers; trim tool outputs to relevant fields; "case facts" block outside summarized history; scratchpad files; crash-recovery manifests loaded by the coordinator.
• Review routing: segment accuracy by doc type/field before trusting aggregates; stratified sampling of high-confidence output; calibrate confidence on labeled sets; claim→source mappings survive synthesis; annotate conflicts with attribution; dates on temporal data.