Reference

Quick-lookup, when you need a rule fast.

The reference pillar is for the moment you forget which tool_choice mode applies, or whether a distractor is a Bigger-Model-vs-Design trap. Concepts answer what is X; scenarios answer how do I build X; reference answers I need the rule right now. Use mid-mock or in the 24 hours before the CCA-F exam.

8 anti-patterns2 cheatsheet2 distractor patterns
01 · Anti-patterns

8 traps that catch most candidates

Severity reflects how often the trap appears on the exam, not how dangerous the bug is in production. A high-severity trap here is one where the distractor is plausible enough to fool a prepared candidate.

PatternDomainSeverityTrapFixShows up in
Subagent context inheritanceD1highTreating a subagent as if it inherits the parent's conversation history. It does not. The parent's tool results, prior turns, and system prompt are invisible until you pass them explicitly.Pass everything the subagent needs in the task prompt: the goal, relevant facts, allowed tools, and return format. Treat each subagent invocation as a fresh session.multi-agent-research-system
18-tool degradation cliffD2mediumStacking 20+ tools on a single agent on the assumption that more capability is better. Selection accuracy degrades sharply past roughly 18 tools as descriptions blur in attention.Cap tools per agent at ~18, or fan out to subagents with narrow tool whitelists. Each subagent does one job well; the coordinator routes.agentic-tool-design
Bigger model for a design problemCrosshighDistractor offers a more capable model, more tokens, or higher temperature. The behavior issue is actually system design: weak tool descriptions, missing few-shot, undecomposed task.Fix the architecture first. Improve tool descriptions, add canonical examples, decompose the task, structure the prompt. Only escalate model after design is sound.Day-of distractor patterns
Prompt-tuned safety guardrailsD4 / D5highRelying on a prompt instruction ("never do X") to enforce a compliance, financial, or security guarantee. Prompt enforcement lands around 70% - fine for ergonomics, fatal for stakes.Use deterministic SDK hooks (PreToolUse / PostToolUse) for hard guarantees. Hooks land at ~100%. Prompts handle preference; hooks handle policy.Day-of distractor patterns
Progressive summarization losing case factsD5highAuto-summarizing the entire conversation as it grows. Critical case-facts (account IDs, prior decisions, user constraints) compress away and the agent contradicts itself two turns later.Pin a case-facts block that never gets summarized. Summarize chat history only. Keep the case-facts block at the start of every turn.case-facts-block
Lost-in-the-middleD5mediumAssuming the model attends evenly across a long context. Mid-context content is reliably under-weighted; key instructions placed in the middle get ignored.Place critical instructions at the start and end. Repeat the most important rule at the end of the system prompt. Keep working context tight.context-window
tool_use output without schema retryD4mediumTrusting tool_use output to match your schema on the first call. JSON validation fails silently downstream and the agent appears to succeed.Validate against a JSON schema. On failure, retry with the validation error in the next user message. Bound retries (2-3 max) and escalate.structured-outputs
CLAUDE.md hierarchy confusionD3lowEditing ~/.claude/CLAUDE.md and wondering why project-specific instructions don't apply, or vice versa. Precedence is layered, not flat.Project-level overrides home. Subdirectory CLAUDE.md files cascade and override their parent. Check the resolved hierarchy with /memory before editing the wrong file.claude-md-hierarchy
02 · Distractor patterns

The two patterns that decide the most exam questions

Once you internalise these two, you will recognise the answer-shape on a meaningful chunk of the mock from the distractor pattern alone. Both are documented from the public CCA-F study material plus the practice-bank misconception analysis surfaced by the adaptive plan.

Model vs Design

Distractor shape. Suggests a bigger model, more tokens, or a higher temperature.

Correct shape. Fixes system design: tool descriptions, few-shot examples, task decomposition, prompt structure.

Rule. Architecture first. Only escalate model after design is sound.

Prompt vs Hook

Distractor shape. Suggests adding instructions to the system prompt to enforce a safety / compliance / financial constraint.

Correct shape. Uses deterministic SDK hooks (PreToolUse / PostToolUse) - ~100% enforcement vs ~70% for prompt tuning.

Rule. If being wrong has real cost, the right answer involves a hook.

03 · Cheatsheets

Per-rule single-page references

Each cheatsheet is a one-page rule lookup with selection table, pitfalls, and drill links. We add cheatsheets when exam-takers ask for them - not preemptively.

Want a cheatsheet on a specific topic? Open an issue on the GitHub repo with the rule and your use-case.

04 · Printable posters

Three downloadable cheatsheet posters

Each poster is a 1080x1920 painterly SVG that downloads as a 2160x3840 PNG. Print on A4 portrait, pin above your desk for the week before the exam, or save and reread 24 hours out.

05 · How to use this pillar

When to come here vs Concepts vs Scenarios

  • Concepts answer what is X. Read once.
  • Scenarios answer how do I build X. Read once or twice.
  • Reference (this pillar) answers I need the rule right now. Open mid-mock, mid-build, or in the 24 hours before sitting.

The anti-patterns table above is the fastest pre-exam refresher we ship. Read it the morning of the exam and one of the eight will surface in your first ten questions.

FAQ

6 questions about the reference catalog

Every Q is phrased as a real Google search query. Answers cite the same evidence-tagged sources used elsewhere on the site.

What's in the reference section?
Three asset types. Cheatsheets - single-page decision flows for primitives where the choice between options is exam-tested (tool_choice auto/any/none, stop_reason branching, MCP routing). Anti-pattern catalog - 20 patterns historically seen as wrong-answer distractors. Posters - 3 downloadable painterly cheatsheets for offline study.
Why is there an anti-pattern catalog?
Because CCA-F distractors aren't random - they reflect mistakes engineers actually make in production. The catalog enumerates 20 of them (e.g., 'using tool_choice required when you need optional tool calls', 'treating stop_reason end_turn and max_tokens identically'). Memorizing the anti-patterns is faster than memorizing 'correct' patterns because the exam's distractors converge on a finite set of trap families.
How are anti-patterns graded for severity?
Three tiers based on observed exam frequency in beta cohort debriefs. High severity (red badge) - patterns flagged in 50%+ of debriefs as 'I picked this and got it wrong'. Medium (yellow) - patterns flagged in 20-50%. Low (gray) - patterns documented in Anthropic best-practice docs but rarely tested. Spend study time proportionally - high-severity first.
What are CCA-F distractor patterns?
The exam writes wrong answers that look right by mirroring real-world bad practices. Common families: (1) 'optimistic stop_reason' - assuming the loop is done when stop_reason is end_turn but a tool result is pending. (2) 'over-eager tool_choice' - using tool_choice required when a model decision is needed. (3) 'PrePrompt vs system prompt' confusion. (4) Hooks misuse (PreToolUse policy gates vs PostToolUse audit logs). The catalog has all 45.
Are the painterly posters worth printing?
Yes if you study offline. The 3 posters compress the highest-frequency exam material - 5-domain weight map, anti-pattern severity quick-reference, tool_choice decision tree - into single-glance painterly compositions. Each is 1024×1280 PNG, free to download, no signup required.
How often is the reference section updated?
Anti-patterns are reviewed weekly during beta cohort intake - new debriefs add to or update the catalog. Cheatsheets refresh on the 60-day cadence or whenever Anthropic ships a new primitive (e.g., new tool_choice options). Posters are versioned and updated quarterly.