The recurring distractor pattern
Across all 12 official sample questions, wrong answers cluster into the same shapes:
relying on the model's probabilistic compliance (prompt wording, few-shot examples, self-reported confidence)
when the situation demands a deterministic/programmatic guarantee — or the reverse, over-engineering infrastructure (custom classifiers, routing layers) when a small prompt/description fix is the proportionate first step. Learn to ask: "does this need a guarantee, or just better guidance?"
| Task | One-line summary |
|---|
| 1.1 | Agentic loop: check stop_reason (tool_use vs end_turn), append tool results, loop. Never parse text or use iteration caps as the stopping condition. |
| 1.2 | Hub-and-spoke: coordinator routes all inter-subagent comms; subagents don't share memory automatically. |
| 1.3 | Task tool spawns subagents (needs allowedTools: ["Task"]); pass context explicitly; parallel = multiple Task calls in one turn. |
| 1.4 | Programmatic prerequisites (hooks/gates) for deterministic ordering; prompts alone have non-zero failure rate. |
| 1.5 | Hooks (PostToolUse, call interception) for guaranteed compliance — normalize data, block policy violations. |
| 1.6 | Prompt chaining (fixed sequential) vs. dynamic decomposition (adaptive) — pick based on predictability of the task. |
| 1.7 | Session mgmt: --resume <name> to continue; fork_session to branch; fresh session + summary beats resuming stale tool results. |
| Task | One-line summary |
|---|
| 2.1 | Tool descriptions are THE tool-selection mechanism. Minimal/overlapping descriptions → misrouting. Fix descriptions before adding infrastructure. |
| 2.2 | Structured MCP errors: isError, errorCategory (transient/validation/permission/business), isRetryable. Generic "failed" messages block recovery. |
| 2.3 | Fewer tools per agent (4–5, not 18) improves selection reliability. Scope tools to role. tool_choice: auto / any / forced. |
| 2.4 | .mcp.json (project, shared) vs ~/.claude.json (personal). Env-var expansion for secrets. MCP resources expose catalogs to cut exploratory calls. |
| 2.5 | Grep = content search, Glob = path patterns, Edit needs unique anchor text (fallback: Read+Write). |
| Task | One-line summary |
|---|
| 3.1 | CLAUDE.md hierarchy: user (~/.claude, not shared) → project (root/.claude, shared via VCS) → directory. @import for modularity. |
| 3.2 | Commands: .claude/commands/ (project) vs ~/.claude/commands/ (personal). Skills: context: fork isolates output, allowed-tools restricts, argument-hint prompts for args. |
| 3.3 | .claude/rules/ with YAML paths: glob-scoping beats directory-level CLAUDE.md when conventions cut across directories (e.g. all *.test.tsx). |
| 3.4 | Plan mode = architectural/multi-file/ambiguous scope. Direct execution = single-file, well-understood change. |
| 3.5 | Concrete I/O examples, test-first iteration, "interview pattern," batch interacting issues vs. sequential independent ones. |
| 3.6 | -p/--print for non-interactive CI. --output-format json + --json-schema for machine-parseable findings. Independent review instance > self-review. |
| Task | One-line summary |
|---|
| 4.1 | Explicit categorical criteria beat vague instructions ("be conservative") for precision/false-positive reduction. |
| 4.2 | Few-shot examples: best tool for consistent format + generalizing judgment to ambiguous/novel cases. |
| 4.3 | tool_use + JSON schema eliminates syntax errors (not semantic ones). Nullable fields prevent fabrication. tool_choice forces a specific tool first. |
| 4.4 | Retry-with-error-feedback works for format errors; doesn't work when info is simply absent from source. |
| 4.5 | Message Batches API: 50% cheaper, ≤24h window, no SLA, no mid-request tool calls. Use for non-blocking workloads only; custom_id correlates. |
| 4.6 | Independent review instance (no generator's reasoning context) beats self-review. Split large reviews: per-file pass + cross-file integration pass. |
| Task | One-line summary |
|---|
| 5.1 | "Lost in the middle": keep a persistent "case facts" block; trim verbose tool output to relevant fields; put key summaries first/last. |
| 5.2 | Escalate on: explicit customer request, policy gap/exception, inability to progress, multiple ambiguous matches. NOT on sentiment or self-reported confidence. |
| 5.3 | Structured error context (failure type, attempted query, partial results) → coordinator recovery. Never silently suppress or hard-terminate on one failure. |
| 5.4 | Scratchpad files + subagent delegation counteract context degradation in long exploration. /compact when verbose. Manifests for crash recovery. |
| 5.5 | Aggregate accuracy hides per-segment failure. Stratified sampling + field-level confidence calibrated on labeled data, not vibes. |
| 5.6 | Preserve claim→source mappings through synthesis. Annotate conflicts (don't arbitrarily pick one). Require dates for temporal data. |
Fine-tuning · API auth/billing/rate limits · specific programming language/framework internals · MCP server hosting/infra · Claude's internal architecture/training/RLHF · embeddings/vector DBs · computer use · vision · streaming API internals · prompt caching internals · tokenization algorithms · cloud provider (AWS/GCP/Azure) specifics · performance benchmarking/model comparison metrics.
Straight from the official guide's own appendix — if a term here doesn't ring a bell, that's a gap to close before exam day.
| Area | What to know |
|---|
| Claude Agent SDK | Agent definitions, agentic loops, stop_reason handling, hooks (PostToolUse, tool call interception), subagent spawning via Task, allowedTools |
| MCP | Servers, tools, resources, isError flag, tool descriptions, tool distribution, .mcp.json, environment variable expansion |
| Claude Code | CLAUDE.md hierarchy, .claude/rules/, .claude/commands/, .claude/skills/ (context: fork, allowed-tools, argument-hint), plan mode, direct execution, /memory, /compact, --resume, fork_session, Explore subagent |
| Claude Code CLI | -p/--print, --output-format json, --json-schema |
| Claude API | tool_use with JSON schemas, tool_choice (auto/any/forced), stop_reason values, max_tokens, system prompts |
| Message Batches API | 50% cost savings, 24h window, custom_id correlation, polling, no multi-turn tool calling |
| JSON Schema / Pydantic | Required vs. optional, enums, nullable fields, "other"+detail pattern, strict mode, semantic vs. syntax validation |
| Built-in tools | Read, Write, Edit, Bash, Grep, Glob — purpose and selection criteria for each |
| Prompting | Few-shot for ambiguous scenarios and format consistency, prompt chaining, explicit criteria |
| Context & reliability | Token budgets, progressive summarization, lost-in-the-middle, scratchpad files, session isolation, confidence calibration, stratified sampling |