Trap. Assuming a subagent sees the parent's conversation. It does not: it starts with only what the orchestrator passes in. Fix. Pass an explicit brief (goal, constraints, identifiers) and require a structured return. Treat the boundary as an API, not a shared memory.
Quick-lookup, when you need a rule fast.
Concepts answer what is X. Scenarios answer how do I build X. Reference answers I need the rule right now: mid-mock, mid-build, or in the 24 hours before you sit. Most of it applies across all four Claude certifications.
01 · Anti-patterns
8 traps that catch prepared candidates
Severity is how often the trap appears on the exam, not how dangerous the bug is in production. Read this the morning of the exam; one of the eight will surface in your first ten questions.
Trap. Reading end_turn as 'task complete'. max_tokens and tool_use both look like a finished message to code that only checks for text. Fix. Branch on stop_reason first. tool_use → run the tool and continue; max_tokens → continue the turn; end_turn → the model is done.
Trap. Exposing 30+ tools at once. Selection quality falls off around 18–20; the model starts picking plausible-but-wrong tools. Fix. Group tools by task and route: expose the 5–8 relevant to the current step, or put a router in front.
Trap. Coding escalation as an exception that kills the run. Escalation pauses the loop for a human; it does not end it. Fix. Model escalation as a state: persist the session, notify, and resume with the human's decision as a new tool result.
Trap. Summarising history into prose and losing order IDs, file paths and decision records. The agent then re-asks or hallucinates them. Fix. Compaction keeps: goal, decisions, open items, every identifier. Drops: verbatim tool output and pleasantries.
Trap. Putting a timestamp or user name before the cache breakpoint. Every request misses; the bill doubles quietly. Fix. Static first: system prompt and tool definitions in a fixed order, breakpoint, then anything that changes per call.
Trap. Hundreds of lines of preferences in one file. Rules conflict, the important ones get lost, and every session pays the token cost. Fix. Short, layered files: user-level defaults, project-level rules that override them, hooks for anything that must be mechanical.
Trap. Re-sending the same malformed tool call three times and hoping. The model rarely changes course without new information. Fix. Return the validation error as the tool result with the schema expected. One informed retry, then escalate.
Two shapes decide most exam questions
Once these are internalised you will recognise the answer shape from the distractor alone.
The deterministic-vs-probabilistic swap
An answer that asks the model to do something code should do: count tokens, enforce a limit, validate a schema, decide when to escalate.
Code owns thresholds, validation and routing. The model owns judgement inside the boundaries code sets.
The 'more context' reflex
An answer that fixes a failure by adding more to the prompt: more history, more tools, more examples, bigger window.
An answer that isolates: a subagent with a narrow brief, a compaction step, a smaller routed tool set.
One-page rule lookups
Selection table, pitfalls and drill links. Added when exam-takers ask.
tool_choice: auto, any, or forcedDecision cheatsheet for tool_choice - auto, any, or forced. When each mode applies and the pitfalls to avoid.D2Hooks vs CLAUDE.md Instructions: Advisory vs Guaranteed ConstraintsDecision cheatsheet for when to enforce a constraint with a hook vs. an instruction in CLAUDE.md - advisory context, guaranteed execution, and the CCA-F exam decision rule.D1 · Agentic Architectures
6 questions about the reference catalog
What is the difference between Concepts, Scenarios and Reference?
Concepts explain a single primitive. Scenarios show how several combine in a production system. Reference is quick-lookup: rules, traps and fixes you need mid-mock or mid-build, without the teaching.
Are the anti-patterns specific to CCA-F?
They are written against the CCA-F blueprint, but six of the eight apply unchanged to Architect – Professional and Developer – Foundations. Associate candidates need the two distractor patterns most.
How is severity decided?
By how often the trap appears in scored questions across our mock bank, not by production impact. High means it surfaces in most 30-question mocks.
Can I download the cheatsheets?
The posters are downloadable PNGs sized for A4. Cheatsheets are web pages so they can link to drills; print from the browser if you want paper.
When do you add a new cheatsheet?
When exam-takers ask for one and the rule is stable in Anthropic's documentation. Open an issue with the rule and your use case.
How often is Reference reviewed?
On the same 60-day cadence as the rest of the site, plus whenever Anthropic publishes new guidance that changes a rule.