Reference · print this · read before the exam

Decision Cheat-Sheet

Every "when you see X, pick Y" rule in one place. The exam is trade-off judgment — these are the trade-offs.

The two master heuristics 1. Guarantee needed → deterministic code (hook, prerequisite gate, forced order). Prompts/few-shot are probabilistic — never the answer when the stem says "guarantee," "must," "reliability," "financial."

2. Quality/calibration problem → proportionate first step = better descriptions / explicit criteria / few-shot, BEFORE infrastructure (classifiers, routing layers, sentiment, self-reported confidence). Watch for "first step" / "most effective" framing.

Domain 1 — Agentic Architecture

Loop control	`stop_reason`: continue on `tool_use`, stop on `end_turn`. NOT prose/iteration-cap/text-presence.
Tool result ignored next turn	You didn't append it to history (stateless API).
Coverage gaps + subagents all succeeded	Coordinator decomposition too narrow (upstream, not downstream).
Subagent ignores prior findings	Context isolated — pass it explicitly in the prompt.
Parallel subagents	Multiple `Task` calls in ONE response. Delegating at all → `allowedTools` has `"Task"`.
Must verify before financial op	Programmatic prerequisite gate (deterministic), not prompt.
Normalize heterogeneous tool data	`PostToolUse` hook.
Block over-threshold action	Tool-call interception hook.
Human handoff (no transcript access)	Structured summary: ID, root cause, amount, recommended action.
Predictable multi-aspect work	Prompt chaining. Open-ended discovery → dynamic decomposition.
Resume vs fresh	Prior valid → resume; stale tool results → fresh + summary; divergent branches → fork.

Domain 2 — Tool Design & MCP

Similar tools confused	First step: expand descriptions (purpose, inputs, examples, boundaries).
Generic tool misused	Split into purpose-specific tools; or rename+re-describe.
Tool failure	Structured error: category + `isRetryable` + readable msg.
Policy violation error	Business, not retryable, explain. Timeout → transient, retry.
Zero matches	Success with empty set ≠ access failure.
18 tools, poor selection	Scope per role (least privilege); scoped cross-role tool for common case.
Model returns text not tool	`tool_choice: "any"`. Specific tool first → forced selection.
Shared server + secret	`.mcp.json` + `${ENV}`. Personal → `~/.claude.json`.
Agent explores to find data	Expose an MCP resource (catalog).
Content search / filename / non-unique Edit	Grep / Glob / Read+Write.

Domain 3 — Claude Code Config

Teammate missing shared instructions	They're in user-level config (not shared) → move to project.
Convention across scattered files	Glob rule in `.claude/rules/` (not per-dir CLAUDE.md).
Team-wide slash command	`.claude/commands/` (config.json array doesn't exist).
Verbose skill clutters chat	`context: fork`. Restrict tools → `allowed-tools`. Missing args → `argument-hint`.
Always-on standard vs run-sometimes	CLAUDE.md vs skill.
Large/architectural/multi-approach	Plan mode (now, if complexity is stated). Simple scoped → direct.
Verbose discovery exhausts context	Explore subagent (returns summary).
CI job hangs	`-p`/`--print`. Structured findings → `--output-format json`+`--json-schema`.
Objective review of generated code	Independent instance (not self-review, not bigger window).
Interacting vs independent fixes	One message vs sequential.

Domain 4 — Prompt Engineering & Output

Too many false positives	Explicit categorical criteria (not "be conservative"/confidence filter).
Inconsistent format / ambiguous judgment	2–4 few-shot examples showing the boundary.
Guaranteed schema-valid output	tool_use + JSON schema (kills syntax errors, NOT semantic).
Sums wrong / value in wrong field	Semantic — validation flow (calculated vs stated), not schema.
Model fabricates absent field	Make field nullable/optional.
Retry works?	Yes for format/structure; NO when info absent from source.
Latency-tolerant / overnight	Batches API (50% off, no SLA, no multi-turn tools, `custom_id`). Blocking → sync.
Batch failures	Resubmit by `custom_id`, chunk oversized.
Big review uneven/contradictory	Per-file + cross-file passes (window size irrelevant).

Domain 5 — Context & Reliability

Facts going vague in long session	Case-facts block in every prompt.
Long input, middle dropped	Key summary first + section headers.
Verbose tool output	Trim to relevant fields.
Codebase session degrading	Scratchpad files + subagent delegation + summarize between phases; crash → manifest.
Escalate?	Human requested (now), policy gap/ambiguous, can't progress. NOT sentiment / self-confidence.
Multiple customer matches	Ask for more identifiers; don't guess.
Subagent failure propagation	Structured context; local recovery first; partials up; annotate gaps; never silent-suppress or kill workflow.
Automate high-confidence extraction?	First segment accuracy by doc type + field (aggregate hides gaps).
Route review attention	Field-level confidence CALIBRATED on labeled set + stratified sampling. (Calibrated ≠ self-reported.)
Conflicting credible sources	Present both w/ attribution; check dates; don't pick/average/drop.

Distractor archetypes — reject on sight (unless the stem truly calls for them) Train a classifier / build a routing layer (over-engineering). Rely on sentiment or self-reported confidence (uncalibrated). Bigger context window fixes attention (it doesn't). Blame a downstream agent (look upstream). Prompt/few-shot when a guarantee is required (use a hook). A fabricated feature (--batch, CLAUDE_HEADLESS, config.json commands array).