Common CCA-F Exam Mistakes to Avoid (2026)

Most CCA-F questions have a plausible wrong answer sitting right next to the right one, and the wrong answer is usually a specific, nameable mistake. Learn the mistakes and you learn to spot the trap. This guide collects the ones that cost the most points, mapped to the domains where they bite, so you walk in already immune to the exam's favorite distractors.

For the full picture, pair this with how to pass on your first attempt and the domains breakdown.

Mistake 1: Reaching for Multi-Agent When One Agent Would Do

The single most-tested judgment error. Multi-agent orchestration is powerful, so candidates over-select it. But Anthropic's own guidance, and the exam's, is to start with the simplest architecture that works and escalate only when the task exceeds one context window, needs real parallelism, or benefits from adversarial separation. When a scenario is solvable by one well-prompted agent, the multi-agent option is the distractor. Read the orchestration deep dive for when delegation actually earns its cost.

Mistake 2: Putting a Hard Requirement in Advisory Prose

A requirement that says "must," "never," or "block" is not satisfied by adding a line to CLAUDE.md, because CLAUDE.md is probabilistic and degrades as context fills. Hard guarantees belong in hooks (a PreToolUse gate that exits 2) or permission rules (deny). The exam pairs a security or correctness "must" with "add it to CLAUDE.md" precisely to see if you know the difference between suggesting and enforcing.

Mistake 3: The Inverse, Over-Engineering a Soft Preference

The trap also runs the other way. A style preference ("prefer functional patterns in this repo") does not need a hook; it needs a CLAUDE.md line. Building enforcement machinery for something advisory is as wrong as under-enforcing something hard. Match the mechanism's strength to the requirement's strength.

Mistake 4: Misranking Structured-Output Methods

When a scenario needs guaranteed-parseable JSON, "instruct the model to please return JSON" is never the best answer. The reliability ranking is tool-use schemas first, then prefilling, then format-plus-example, and unconstrained requests last. Picking a fragile method when a schema-constrained one is offered is a giveaway wrong answer in the structured output domain.

Mistake 5: Confusing Skills, Slash Commands, and MCP

Three different mechanisms, three different triggers:

• A skill is loaded by the agent on demand.
• A slash command is triggered by the user typing it.
• MCP connects the agent to external systems.

"Use a skill to connect to the database" (that is MCP) and "use MCP to teach the refund procedure" (that is a skill) are the exam's favorite swaps. The skills and slash commands guide draws the lines.

Mistake 6: Getting Settings Precedence Backwards

Enterprise managed policy outranks project settings, which outrank user settings. A question where "a developer's personal config loosens a company security rule" is testing whether you know the org's deny wins. Candidates who assume the more local (personal) setting wins because it is "closer" get it backwards. Precedence flows from organizational authority down, not from local up.

Mistake 7: Writing Tool Errors for Humans, Not the Model

In tool design, an error message is a prompt back to the model. "Error 422" teaches the agent nothing and it gives up or loops; "Invalid date format for start_date, expected YYYY-MM-DD" tells it how to retry. Candidates from a traditional API background write errors for a developer reading a log and miss that the reader here is a model deciding its next move. Covered in the tool design guide.

Mistake 8: Ignoring the Context Budget

Treating the context window as infinite shows up as several wrong answers: bloated CLAUDE.md files that slow every session, tools that return whole records when a field would do, and reaching for a bigger context instead of a subagent when a job is noisy. The exam's Context Management domain rewards treating context as a budget you spend deliberately.

Mistake 9: Choosing the Wrong MCP Transport

stdio for local, subprocess-spawned servers; Streamable HTTP for remote, load-balanced deployments; HTTP+SSE is deprecated for new work. A scenario describing a multi-tenant remote server whose "right" answer is stdio, or a local Claude Desktop tool whose answer is a load balancer, is testing this mapping. The MCP architecture guide has the transport table.

Mistake 10: Exam-Mechanics Errors

Not every mistake is conceptual:

• Rushing scenario stems. The question is often in the last sentence; read it before the setup so you know what to look for.
• Second-guessing. On scenario exams, your first well-reasoned answer is usually right; change only with a concrete reason.
• Leaving questions blank. There is no penalty for wrong answers, so flag and return, but never leave one empty.
• Studying evenly across uneven domains. Orchestration is 27% and context is 15%; equal study time misallocates your hours.

Key Takeaways

• Default to the simplest architecture; multi-agent is over-selected
• Match mechanism strength to requirement strength: hooks/permissions for "must," CLAUDE.md for "prefer"
• Rank structured-output methods correctly; schema-constrained tool use wins for reliability
• Keep skills, slash commands, and MCP straight by who triggers and what they do
• Precedence flows from enterprise authority down; org deny rules are absolute
• Treat context and tool results as a budget; write tool errors for the model to act on

The surest way to internalize these is to meet them in practice questions. Preporato's CCA-F practice tests drill every trap above across 390 explained questions, with the flashcard deck locking in the tables the mechanics questions test.

Sources:

• Anthropic: Building Effective Agents
• Anthropic Academy

Last updated: July 10, 2026