Skip to content

CPES - Exam Strategy

The Specialist exam tests judgment more than recall. Many questions show two prompt variants and ask which is better, or show a failing prompt and ask which intervention would help most. Your job is to recognize the technique each option represents and judge its fit.


3-Phase Preparation

Phase 1 - Read (Week 1)

Read Anthropic's prompt engineering guide twice. Make a one-page cheat sheet of techniques with examples. The exam vocabulary tracks the docs precisely.

Phase 2 - Practice (Weeks 2-3)

Iterate on a real task with measured evals. You should ship at least 5 prompt versions and know which technique drove which improvement.

Phase 3 - Drill (Week 4)

Run timed scenarios. Re-read the fact sheet. Sleep before the exam.


Time Management

Assume 55 questions in 90 minutes. ~1m 35s per question with a small buffer.

Time Elapsed Question #
25 min 16
50 min 32
75 min 48
85 min 55 (done, review)

First pass: answer fast, flag the slow ones. Second pass: spend buffer on flags.


Reading Prompt Snippets Fast

Train yourself to spot:

  • XML tag usage and consistency
  • Where the question is placed (top, middle, bottom)
  • Whether examples are in the prompt
  • Whether the system field is used at all
  • Whether cache_control appears
  • Whether tool_choice is forced for structured output
  • Whether prefill is used

Most "what is the best prompt" questions test one of these axes. Identify the axis first.


Common Question Patterns

"Which prompt is more effective?"

Look for:

  • Better use of XML structure
  • Explicit format and constraint specification
  • Examples present and well-chosen
  • Question position (top + bottom for long context)
  • System prompt used appropriately

"What is the most effective change?"

Identify the failure mode the prompt currently exhibits, then match the intervention:

  • Wrong format -> examples or forced tool choice
  • Wrong tone -> system prompt with persona
  • Bad reasoning -> chain of thought or extended thinking
  • Verbose -> length constraint
  • Lost in long context -> question at top and bottom; XML tags
  • Inconsistent across calls -> few-shot anchoring
  • Refuses inappropriately -> rephrase task; system prompt clarification

"Why is this prompt failing?"

Look for ambiguity, missing format spec, missing examples, dynamic content in cached regions, role prompts used as safety controls.


Answer Selection Heuristics

  • Prefer documented Anthropic patterns over clever alternatives
  • Prefer the simpler intervention that addresses the named failure
  • Prefer techniques that compose (XML + examples + CoT) when the question allows
  • Prefer extended thinking over prompt-level CoT when the model supports it AND you do not need visible reasoning
  • Prefer forced tool choice for strict structured outputs
  • Prefer examples over instructions for format requirements

Domain-Specific Tactics

Prompt Fundamentals

Look for clarity, specificity, and ordering. The best prompt is usually the most explicit one within reason.

System Prompts

System prompts persist; user prompts vary. Constraints and persona belong in system. The actual task belongs in user.

Chain of Thought

Three flavors: implicit ("think step by step"), structured (numbered steps), tagged (<thinking>/<answer>). All work; tagged is most parseable.

XML Tags

Use them. Be consistent. Tag names should be meaningful for the domain.

Few-Shot

Quality of examples beats quantity. Diversity beats redundancy. Include edge cases.

Evaluation

Eval datasets must have gold or rubric. LLM judges must be calibrated. A/B test deliberately.

Prompt Caching

cache_control on content blocks. Static content first. Dynamic content last. Verify with usage fields.


Counter-Intuitive Facts to Internalize

  • Adding "be helpful and concise" alone often does little. Specify tokens or sentences.
  • Long instructions can outperform short ones on hard tasks.
  • Few-shot helps even when the task is well-described.
  • Extended thinking can underperform a well-crafted prompt with no thinking on simple tasks (cost without benefit).
  • Prefill is one of the most underused techniques.
  • Roles do not bypass refusals.

The 24 Hours Before

  • Re-read fact sheet
  • Re-read this strategy
  • Sleep
  • Have ID and stable internet ready

During the Exam

  • Read every prompt snippet fully; do not skim
  • Identify the technique axis the question targets
  • Eliminate two answers fast
  • Choose the answer most aligned with Anthropic's documented patterns

If You Fail

Score reports show your weakest domain. Spend two weeks rebuilding it with hands-on exercises. Re-sit. Specialist exams reward depth in the craft, which compounds with practice.