I analysed the latest 35,000 answers from the CCAR-F mock exams to find where people struggle. If you are studying to get certified, this might help you decide where to spend your time.
About the data
These are my questions, not Anthropic's. Everyone who answered them chose to study with me.
The weakest domain is the one you use every day
Of the five exam domains, people struggle most with Claude Code Configuration & Workflows.
TL;DR
Using Claude Code every day is no reason to study less for it. The exam asks about the parts you rarely use: CI/CD, how the GitHub Action behaves and the CLAUDE.md hierarchy.
Hardest task statement: iterative refinement, 43% correct
Two traps cause the low score.
The code must meet a time limit, and two attempts to make it faster have failed. 54% chose "design it in plan mode first". 20% chose the right answer: write a test that fails when the code exceeds the time limit, and run that test on every attempt. Plan mode is for comparing approaches before a large change. A test checks each attempt against a measurable target.
Three bugs all affect the same output. 59% chose "fix the root one first, then send the others one at a time". The exam guide says to send interacting issues in one detailed message, and independent issues one at a time.
How people answered the two traps
Code misses a time limit after two failed attempts
- 54%Design it in plan mode firstMost picked
- 20%A failing test for the time limit, run on every attemptCorrect
- 26%The other two answers
Three bugs affect the same output
- 59%Fix the root one first, then send the others one at a timeMost picked
- 41%Every other answer, the correct one included
People choose hooks too often
Most people know that hooks enforce a rule more reliably than prompt instructions. When a rule must apply to every call, 93% chose correctly.
In another question, a miss costs little and the hook adds latency. Only 28% chose the lighter fix: put the rule in the prompt and measure how often the agent misses it. The most common answer kept the hook.
Same tool, two situations
Share who chose the right answer.
- A miss is costlyRight answer: a hook93%
- A miss is cheap, the hook adds latencyRight answer: the prompt28%
The exam expects a hook when a miss is costly, and a prompt instruction when a miss is cheap.
Four more traps that catch a quarter of people or more
- 41%fell for it
Claude writes prose when it should call a tool
41% chose to expand the tool description. 39% chose the right answer.
FixReword the system prompt instruction that told Claude to do the check itself. Prompt wording can override a good tool description. - 1 in 3fell for it
claude-code-actionwith nopromptinputA third of people thought it uses CLAUDE.md as the prompt.
FixIt runs in interactive mode and waits for@claudein a comment. - ~1 in 3fell for it
A downstream agent fills its context window
Almost a third of people moved the long output into a scratchpad file.
FixChange the agent that produces the output, so it returns structured facts and citations instead of its reasoning. - 1 in 4fell for it
An agentic loop follows a fixed script
A quarter of people kept the script and added a step.
FixLet the model choose the next tool from the context.
Wrong answers aren't random
On single-answer questions, 59% of wrong answers go to one distractor. If people guessed at random, each distractor would get about a third.
Where the wrong answers go
Each single-answer question has three distractors.
What people picked
- 59%The most-picked distractor
- 41%The other two combined
Random guessing
- 33%Any one distractor
- 67%The other two combined
Takeaway: when two answers look right, find the detail in the question that rules one of them out.
First attempts and retakes
The average score on a first full mock was 699. The pass mark is 720. 46% passed at the first attempt. Of those who failed, 28% were within 50 points of a pass.
The first full mock, against the pass mark
- Average first score
- 699
- Pass mark
- 720
- Passed first time
- 46%
- Of fails within 50 points (shaded)
- 28%
People who kept practising scored higher. People who took it three times went from 69% correct on their first attempt to 80% on their third.
Three attempts, same people
Share of answers correct, for people who took the mock three times.
- First attempt69%
- Second attempt75%
- Third attempt80%
Where people do well
People score highest on multi-pass review (90%), structured error responses (89%) and few-shot prompting (85%). If your study time is short, spend less of it on these topics.
The strongest task statements
- Multi-pass review90%
- Structured error responses89%
- Few-shot prompting85%
- Iterative refinementThe hardest, for comparison43%