Read the morning of your exam. Every entry below is a pattern that maps directly to a Skills-in bullet in the official guide — print this page, scan it once, then close it.
When a question describes one of these failure modes, the right answer almost always points to the matching pattern.
| If the question describes… | …the answer is |
|---|---|
| Critical sequence violated (e.g., refund before customer verified) | Programmatic prerequisite (hook) |
| Tool selection unreliable for ambiguous requests | Expand tool descriptions first; then 4–6 few-shot examples with reasoning |
| Bad escalation calibration | Explicit criteria + few-shot — not sentiment, not self-reported confidence |
| Multi-concern requests with redundant fetching | Decompose into distinct concerns, investigate in parallel with shared context, synthesize |
| Output gaps vary case-by-case (different things missing each time) | Self-critique step (evaluator-optimizer) — not few-shot |
| Prose has been interpreted differently each iteration | 2–3 concrete input/output examples |
| Critical facts lost to summarization | Case facts block outside the summarized history |
| Self-review misses subtle issues | Independent second instance (no shared reasoning context) |
| CI hangs on interactive input | -p / --print flag |
| Need structured output for CI tooling | --output-format json + --json-schema |
| Team-wide slash command needed | .claude/commands/ in project repo |
| Personal customization without affecting teammates | ~/.claude/skills/ with a different name |
| Conventions span multiple directories | .claude/rules/ with YAML paths: globs |
| Multi-agent topic coverage gap | Coordinator decomposition — not subagent execution |
| Subagent failure | Structured error context: failure type + attempted query + partial results + alternatives |
| Multi-file PR producing inconsistent depth | Per-file passes + separate cross-file integration pass |
| Round-trip overhead from coordinator delegation | Scoped cross-role tool for the common case (e.g., verify_fact) |
| Generic tool being misused outside its specialty | Replace with constrained alternative (fetch_url → load_document) |
| "Lost in the middle" on aggregated input | Key findings at top + explicit section headers |
When a skill exhibits multiple structural issues, the answer almost always combines these three frontmatter keys — not prose instructions.
| Frontmatter | What it solves |
|---|---|
context: fork | Verbose / exploratory output that would pollute the main session |
allowed-tools | Restricting tool access during skill execution (block destructive ops) |
argument-hint | Prompting developers for required parameters when invoked without args |
Telling Claude in the skill body to "ignore prior context" or "don't use Bash" is unreliable — it's prompt-level, probabilistic guidance. The frontmatter keys give declarative, deterministic enforcement. If you see a question listing 2–3 skill problems, look for the answer that combines two or more of the above.
Four cases. The discriminator is whether the customer made an explicit request for a human vs. just expressing frustration.
| Situation | Action |
|---|---|
| Customer explicitly asks for a human ("I want a real person now") | Escalate immediately — no investigation, no clarifying questions |
| Customer frustrated, no explicit human request | Acknowledge frustration, offer resolution. Escalate only if they reiterate. |
| Policy is silent / ambiguous on the request (e.g., competitor price match when policy covers only own-site) | Escalate — policy gap, not just complexity |
| Tool returns multiple matches (e.g., 3 customers named "John Smith") | Ask for an additional identifier — don't heuristic-pick, don't escalate |
Match the API to the latency requirement, not the cost savings.
50% cost savings · ≤24h window · no SLA · custom_id correlates request/response pairs · no multi-turn tool use
Reject any answer involving these — they're the wrong-answer signatures the exam keeps using.
CLAUDE_HEADLESS, --batch, .claude/config.json — none exist (verify against current Claude Code docs)get_customer_with_orders) when the answer is "prompt Claude to batch tool requests per turn"High-frequency patterns by domain. Each links to its full study guide (when available).
stop_reason: "tool_use" continues, "end_turn" stopsTask must be in allowedToolsPostToolUse normalizes; PreToolUse enforces complianceisRetryabletool_choice: "auto" / "any" / forced.mcp.json (project) vs ~/.claude.json (user); ${ENV_VAR} for secrets.claude/rules/*.md with YAML paths: for cross-directory conventions.claude/commands/ (project) vs ~/.claude/commands/ (user)context: fork, allowed-tools, argument-hint-p prevents hangs · --output-format json + --json-schematool_use + JSON schema fixes syntax — not semantic correctnesscalculated_total vs stated_total; detected_pattern for FP analysisSource: every entry traces to a Skills-in bullet in the official CCA Foundations Exam Guide. Drill on the practice exam →