Claude Certification Blog
Prompt engineering on the Claude exams: repair, not craft
Prompt engineering is named on every Claude certification track and is the largest subject on none of them — because the exams test it as a repair skill rather than as a craft.
Prompt engineering appears on all four Claude certification exams, and on none of them is it the heaviest material. CCAO-F gives it three objectives inside a 14% domain. CCAR-F gives it two inside a 20% one. CCAR-P names three separate objectives — system prompts and templates, technique, and reuse. CCDV-F publishes it as a 4.6% skill. Candidates who arrive expecting a prompt-engineering exam find something narrower: a scenario names a failure and asks which single change addresses it.
Named on all four tracks
| Track | Where it sits | What the objectives cover |
|---|---|---|
| CCAO-F | Domain 1, 14% | Writing, iterating, and adapting to task type |
| CCAR-F | Domain 4, 20% | Explicit criteria, and few-shot for consistency |
| CCDV-F | Skill, 4.6% | Instruction, examples and structure, applied singly |
| CCAR-P | Domain 2, three objectives | System prompts, technique, and reuse strategy |
The domain percentages above cover more than prompting — CCAR-F’s 20% also holds structured output, batching and multi-pass review — so the prompting share is smaller than the number suggests on three of the four. Only CCDV-F publishes the isolated figure, and it is 4.6%. Full weights for every track are in the domain weights post.
Diagnosis, not craft
No track asks you to write the best prompt for a task. What they ask is narrower and more answerable: given this specific problem with this output, which one change addresses it. That framing is why a general answer never wins. Every option in a well-built item is a real technique; only one of them answers the failure described.
Criteria beat adjectives
CCAR-F names designing prompts with explicit criteria to improve precision and reduce false positives, and that objective generalises across the programme. An instruction built on a judgment word — serious, relevant, high quality — hands the decision to whoever is reading it, and the results vary because the word does.
Replace the adjective with the conditions that qualify and the same instruction becomes something a reviewer can disagree with and a measurement can track. That is the whole move, and its failure mode is the recurring distractor: asking for the same thing more firmly. A stronger adjective is still an adjective.
State what to do, not only what to avoid
Telling the model what is unacceptable rules out one region and leaves the rest open. “Do not be verbose” permits every brief answer there is, including the several you did not want. The exam-shaped version names the destination as well as the fence.
What examples actually teach
Few-shot prompting is named directly on CCAR-F and on CCAR-P, and both frame it as a consistency tool rather than a quality one. The examinable subtlety is that examples do not illustrate the rule — they define the target, and they teach more than intended.
Two consequences follow. A set drawn only from clear-cut cases quietly defines the task as easy, so the ambiguous cases the examples avoided are the ones that go wrong. And where an example and a stated rule disagree, the example wins — which means a contradicting example is a correction to make, not a rule to state more loudly. Returns also flatten: once the distinctions the task needs are covered, each further example is paid for on every single call and buys nothing you can measure.
Reuse is its own objective
CCAR-P names implementing prompt reuse strategies — caching, modular prompts and Skills — as a separate objective, which no other track does at that level. It is worth noticing because it reframes a prompt as an artefact with a lifecycle rather than a message you send.
The practical judgment is ordering. Caching rewards material that is stable and comes first, so a prompt assembled with its varying parts up front is more expensive than the same content arranged the other way. And a prompt duplicated across five services is five prompts to update when one thing changes, which is the modularity half of the objective. Where that meets cost on the developer track is covered in model selection and optimisation.
The part people skip
CCAO-F names iterating prompts to improve output quality, and the word doing the work is improve. An iteration you cannot show improved anything is a change, and the exams are consistent about the difference: a baseline, one variable, the same evaluation set, and a threshold fixed before you looked at the result.
The threshold ordering is the part that gets skipped and the part that gets tested. Chosen after the measurement, a bar accommodates whatever number appeared and always passes. Chosen before, it can fail — which is the only condition under which passing means anything. The full discipline is in writing evals for Claude prompts, and the diagnosis half in troubleshooting on the Claude exams.
Key takeaways
- Named on four tracks, largest on none. CCDV-F publishes the only isolated figure and it is 4.6%.
- It is examined as repair. A named failure, and the single technique that addresses that failure specifically.
- Conditions beat adjectives. Asking more firmly for the same judgment word is the recurring wrong answer.
- Examples define the target. Where an example and a rule disagree, the example wins, so fix the example.
- Prompts have a lifecycle. Caching rewards stable content first; duplication makes one change into five.
- Fix the bar before you look. A threshold chosen afterwards describes the result rather than judging it.
The technique is easy; picking the right one under time is not
Every option on a well-built prompting item is a technique that works somewhere. Sorting the one that answers this failure from the three that answer other failures is a timed skill, and timed papers are the only place to practise it. Our claude certification study guide covers where they belong in a plan.
Try a full timed paperQuestions
Frequently asked
The follow-up questions people search next.
Is prompt engineering on the Claude certification exams?
On all four, and on none of them is it the largest thing. CCAO-F Domain 1 is 14% and carries three prompting objectives; CCAR-F Domain 4 is 20% and gives prompting two of its objectives; CCAR-P names three inside Domain 2; and CCDV-F publishes Prompt Engineering as a 4.6% skill.
Are the Claude exams prompt-engineering exams?
No, and expecting that is a common miscalibration. The technique is examined as a repair skill: a scenario describes a specific failure and asks which single change addresses it. Craft questions — write the best prompt for this — do not appear.
Which prompting techniques are named in the guides?
CCAR-P names zero-shot, few-shot and chain-of-thought directly in one objective. CCAR-F names few-shot prompting and designing prompts with explicit criteria. CCAO-F names adapting the approach to the task type — analysis, research, drafting, brainstorming.
What is the standard wrong answer on a prompting item?
Asking more firmly for the same thing. Where a prompt produced vague output, the option that repeats the request with stronger wording changes nothing measurable. The one that replaces an adjective with stated conditions changes what can be judged.
Is prompt caching a prompt-engineering topic?
On CCAR-P it sits inside an objective on prompt reuse strategies alongside modular prompts and Skills, so yes. On CCDV-F the same material sits under cost and token management instead. The judgment is identical: stable content first, so reuse is possible.
Keep reading
Related posts
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change — always confirm against the official exam guide before booking.