CCA-F Cheatsheet

The single-page condensed brain dump. Read every day in the final week.

Domains + weights

1. Agentic Architecture & Orchestration — 27%
2. Tool Design & MCP Integration — 18%
3. Claude Code Configuration & Workflows — 20%
4. Prompt Engineering & Structured Output — 20%
5. Context Management & Reliability — 15%

Five agentic patterns

Pattern	Use when
Prompt chaining	Steps known in advance
Routing	Multiple specialized handlers
Parallelization (sectioning / voting)	Independent subtasks OR consensus
Orchestrator-workers	Dynamic decomposition needed
Evaluator-optimizer	Quality > speed; iterative refinement

Default: simplest pattern that works. Workflow > agent unless dynamism required.

Tool design rules

• One tool, one verb.
• Description = teacher's note: "Use when X, not when Y."
• Enums > strings; optional > required; default values reduce hallucination.
• Return short structured output; large blobs → resource URI.
• Annotate destructive / idempotent / read-only.
• Errors are structured (is_error: true, type, message).

Error types

• transient (retry) | permanent (adapt) | validation (model retries) | uncertain (defensive read) | user_declined | unavailable (graceful).

MCP — who decides

• Tools → model.
• Resources → application (host attaches).
• Prompts → user (slash command).

MCP transports

• stdio = local subprocess.
• SSE = legacy remote.
• Streamable HTTP = current remote standard; OAuth 2.1.

Claude Code config precedence

CLI flags > .claude/settings.local.json > .claude/settings.json (shared) > ~/.claude/settings.json (user) > enterprise.

Deny always beats allow. Allow skips confirmation prompt.

CLAUDE.md hierarchy

~/.claude/CLAUDE.md (personal) ← project shared ← project local (gitignored) ← subdir.

Slash command vs Skill vs Subagent

	Slash	Skill	Subagent
Invoked by	User typing /x	Model (auto)	Model (delegate)
Loaded when idle	None	Description only	None
Contents	Prompt template	Prompt + files + scripts	New agent w/ own context

Prompt-engineering ladder (in order)

1. Clear and direct → 2. Few-shot → 3. Let it think → 4. XML tags → 5. Role → 6. Prefill → 7. Chain.

Structured output — three techniques

1. Force tool use (tool_choice: {type: "tool", name}) — most reliable.
2. Prefill { or <output>.
3. Prompt-only — least reliable.

CALM (context mgmt)

• Cache stable prefixes.
• Align prefix structure.
• Limit history.
• Monitor cache hits + tokens.

Compaction strategies

• Progressive summarization | State object | Sliding window | Retrieval-on-demand | Subagent delegation.

Prompt-caching mechanics

• cache_control: {type: "ephemeral"}.
• 5-min TTL default; 1h tier where available.
• Up to 4 breakpoints.
• Cache write ≈ 125% input; cache read ≈ 10%.
• Pays off after ~2 reuses.
• Any prefix change invalidates.

Batch API

• 50% input + output discount.
• Async, ≤24h SLA.
• Use custom_id for mapping.
• Caches inside the batch.
• For interactive flows: do NOT use batch.

Model choice

Want	Model
Best reasoning	Opus 4.7 (claude-opus-4-7)
Balanced production	Sonnet 4.6 (claude-sonnet-4-6)
Cheap + fast	Haiku 4.5 (claude-haiku-4-5-20251001)

Idempotency rules

• Idempotent → retry safe.
• Non-idempotent → defensive read first; require user confirm.

Six scenarios — anchors

1. Support → routing + tools + escalation.
2. Code-gen Claude Code → CLAUDE.md + plan mode + slash commands.
3. Research → orchestrator-workers + subagents + provenance.
4. Dev productivity → MCP heavy + read-only defaults.
5. CI/CD → --print + narrow allow + hooks.
6. Extraction → forced tool use + Batch API + validation loop.

Test-day discipline

• Read each question twice.
• "Obvious" is usually wrong by one trade-off.
• When two answers look right, pick the one that respects principles over conditionals, and least privilege over convenience.
• Flag uncertainties; review at end.

See also

• cca-f-exam-overview
• study-plan
• cca-scenarios
• glossary