CCA Study Tips for Domain 4: Prompt Engineering Mastery

Study Domain 4 (Prompt Engineering & Structured Output, 20% of the CCA-F) by reading the 25 articles in task order over four sessions and, for every scenario, asking which layer the fix belongs to: the prompt, the schema, or the code around the call. That one habit answers most Domain 4 questions. This post gives you the reading order, what to memorise versus reason about, a practice cadence, self-checks, and the mistakes to review twice.

For what the domain covers and why it is weighted at 20%, see CCA Domain 4 Explained. This post is only about how to study it.

What order should I read the Domain 4 articles in?

Task order on the Domain 4 study guide page is right. Suggested sessions:

Session	Articles	Goal
1	Tasks 4.1 and 4.2: “Be Conservative” Means Nothing, One 60% False Positive Category, Labels Alone: 41%. Text Conditions: 72%. Text + Code Examples: 94%., Text Instructions Failed at 64%. Two Examples Reached 91%., Three Diverse Examples Beat Eight Homogeneous Ones, few-shot hallucination, Paired REPORT + SKIP Examples	Rewrite one vague criterion as a verifiable condition; design a three-example set with a boundary case
2	Task 4.3: tool_use Eliminates Structural Errors. Semantic Errors Remain., Forced, Any, Auto, Required Fields on Optional Data, closed enums	Write a schema with one nullable field and one enum, and say which tool_choice mode you would use
3	Task 4.4: Blind Retry: 12%. Error Feedback: 87%., Retrying Absent Data Causes Hallucination, confidence routing, Line Items Don’t Sum to Total	Classify five example failures as format, structural, absent, or ambiguous
4	Tasks 4.5 and 4.6: 50% Cost Savings — But Up to 24 Hours Wait, custom_id, batch errors, expiry, cost optimisation, Always Test on a Sample, Same-Session Review: 0.3 Findings. Independent Instance: 3.7., per-file and cross-file passes, independent instances, the review pipeline	Sketch a batch job with meaningful custom_ids and a resubmit rule; sketch a two-call review

Before session 2, reread the Foundations pieces on Virtual Tools: Hijacking tool_use for Guaranteed JSON and required, nullable, and enum. They are short and make Task 4.3 land.

What should I memorise, and what should I reason about?

Memorise:

• The three tool_choice guarantees: forced tool always calls that tool; any always calls some tool; auto may return text.
• Nullable syntax: "type": ["string", "null"] and remove the field from required.
• Schema constraints override prompt instructions and few-shot examples. If the schema requires a string, no prompt wording produces null.
• The retry taxonomy: format and structural errors converge with specific feedback; absent data diverges (a different value each retry); ambiguous data gets one disambiguation retry then a qualifier.
• Batch shape: roughly half price, up to 24 hours, arbitrary result order, custom_id is the only correlation (custom_id Is Your Only Correlation), no multi-turn tool use, results expire (Results Expire in 29 Days).
• Independent review means a separate call whose context contains only the artifact and the criteria, not the generation prompt.

Reason about:

• Is the criterion verifiable? If a reasonable colleague could disagree about whether it applies, it is vague. Replace adjectives with conditions.
• Do the examples cover the boundary? Diverse examples including a SKIP or negative case, and at least one showing a null field.
• Which layer fixes this? Vagueness is prompt; shape is schema; consistency across fields, retries and cost are code.
• Retry or accept null? Ask whether repeated attempts converge.
• Sync or batch? Latency tolerance in hours and single-turn requests mean batch; anything interactive or tool-calling stays sync.
• Same session or independent? If the reviewer can see why the code was written, findings will be suppressed (Even Design Goals from the Generation Prompt Suppress Review Findings).

Percentages in titles show direction. The exam does not test the numbers.

What practice cadence works for Domain 4?

• After each session, run 15 to 20 Domain 4 practice questions. Before revealing, write down the layer (prompt, schema, code) you think the answer lives in. Track how often you were wrong about the layer; that number should fall to near zero by session 4.
• After session 4, do a 40-question block. Domain 4 distractors are subtle (an option that “adds an instruction” is often plausible-sounding but wrong), so a longer block helps you notice the pattern.
• Review mistakes every few days. The recurring ones are choosing prompt emphasis over a schema change, retrying absent data, and forgetting that auto can return text.
• Final week: the free CCA-F mock exam will include roughly 12 Domain 4 questions among 60 (confirm current weights on the official Anthropic / Pearson VUE page before you register).

Self-check questions before moving on

1. A classifier’s inter-run agreement is under 50%. What is the most likely cause and the fix?
2. All eight few-shot examples are near-identical positive cases. What two things are wrong?
3. warranty_expiry is a required string and the document has no warranty. What does the model do, and what is the fix?
4. You have three extraction schemas and must always get structure. Which tool_choice mode?
5. Three retries for a license field produce MIT, Apache-2.0, then BSD. What does that tell you?
6. A JSON invoice passes the schema, but line items do not sum to the total. Which validation layer catches it?
7. 985 of 1,000 batch requests succeeded. What do you resubmit, and what do you check first? (940 Succeeded, 45 Errored, 15 Expired, Different Errors Need Different Fixes)
8. Asking the same session to “review strictly” finds almost nothing. Why, and what is the fix?

Mistakes worth reviewing twice

• Capitalising an instruction and expecting a structural guarantee.
• Treating a schema-valid output as a correct output.
• Adding more examples of the same kind instead of more diverse ones (Every Example Shows All Fields Populated — So the Model Fabricates Missing Ones).
• Escalating urgency in a retry prompt for data that is not in the source.
• Forcing minimum finding counts in review, which produces invented issues.
• Using confidence to gate output in Domain 4 style without adding the SKIP examples that fix the underlying miscalibration (Track Which Patterns Developers Dismiss — Then Add SKIP Examples).

Next step

Do session 1 today from the Domain 4 study guide page and finish the Domain 4 practice set within the week, tracking which layer each answer lived in. Then read CCA Exam Anti-Patterns for the cross-domain distractors, and sit the free CCA-F mock exam when your accuracy holds.