CPES - 4-Week Practice Plan¶

Daily commitment: 1-1.5 hours weekdays, 2-3 hours one weekend day. Total: ~35 hours.

This plan is hands-on. Every week ends with measured prompt improvements.

Week 1 - Fundamentals and System Prompts¶

Read Anthropic's prompt engineering overview cover to cover
Read notes/01-prompt-fundamentals-and-anatomy.md and notes/02-system-prompts-and-role-prompting.md
Pick one real task you care about (summarization, classification, extraction, code review, support reply)
Write a v1 prompt without fancy techniques. Save it.
Build a small eval dataset (20 items) with gold answers or rubric
Score v1 against the eval set (manual is fine for week 1)
Add a system prompt with explicit role and constraints. Re-score.
Document what changed and by how much in a prompt journal

Deliverable: a baseline + v2 prompt with measured improvement.

Week 2 - Chain of Thought, XML Tags, Structured Output¶

Read notes/03-chain-of-thought-and-extended-thinking.md and notes/04-xml-tags-and-structured-outputs.md
Add a "think step by step" instruction or <thinking> tag. Re-score.
Refactor your prompt to use XML tags for instructions, context, and examples. Re-score.
If your task has structured output, switch to forced tool choice. Re-score.
Try prefill (start the assistant with { for JSON). Compare with forced tool choice.
Try enabling extended thinking (if your task is reasoning-heavy). Compare with prompt-level CoT.

Deliverable: v3-v5 of the prompt with documented technique-by-technique impact.

Week 3 - Few-Shot, Many-Shot, and Iteration¶

Read notes/05-few-shot-and-example-driven-prompting.md and notes/06-prompt-evaluation-and-iteration.md
Add 3 carefully chosen examples spanning easy / typical / edge case. Re-score.
Try 8 examples (many-shot). Compare to 3.
Add one counter-example. Re-score.
Build an automated LLM-as-judge to replace manual scoring
Calibrate the judge against your manual labels (compute agreement)
Run an A/B between two prompt variants on the eval set
Document which technique gave the biggest lift on your task

Deliverable: an automated eval harness + a winning prompt variant.

Week 4 - Caching, Long Context, Review, Exam Prep¶

Read notes/07-prompt-caching-design-patterns.md
Reorder your prompt to maximize cache-friendly structure
Add cache_control at the end of the static prefix
Verify cache hit rate and token cost in the response usage
If your task uses long documents, apply long-context tips: tags, question top-and-bottom, quote-then-answer
Re-read all notes and the fact sheet
Walk through every scenario in scenarios.md under timed conditions (1.5 minutes per question)
Re-read strategy.md the night before
Verify exam logistics on Anthropic's site

Deliverable: a production-ready, cache-optimized prompt with full eval harness.

Hands-On Project Ideas (Pick One)¶

Customer support reply generator: tone-appropriate replies grounded in product docs with citations.
Invoice extractor: structured JSON via forced tool choice; eval against ground-truth invoices.
Code review assistant: identify issues in a diff with severity tagging.
Research summarizer: long-context summarization with page-level citations.
Multi-step lesson planner: chain-of-thought with structured plan output.

Prompt Journal Format¶

Maintain a journal of every change:

v3 -> v4
Date: 2026-04-12
Hypothesis: Adding two counter-examples will reduce false positives.
Change: Added two counter-examples in <bad_output> tags.
Eval delta: +4.2% accuracy, +800 input tokens.
Decision: Ship.

This discipline pays off at exam time when you need to recognize which technique a question is testing.

Red Flags - Do Not Sit the Exam Yet If¶

You cannot list Anthropic's prompting techniques from memory
You have not measured the impact of any technique on a real task
You confuse extended thinking with CoT prompting
You have never structured a prompt with XML tags
You think prompt caching is automatic
You have never used forced tool choice for structured output