Unofficial reference site
A reference guide to Claude Code & agent-building judgment
This site explains the concepts behind building well with Claude Code, the Claude Agent SDK, and the Claude API — CLAUDE.md and hooks, MCP tool design, agentic loops, structured extraction, and more. It's based on a third-party study guide sometimes called "CCAF" prep material, reframed here as a pure explainer: no quizzes, no hidden answers — every idea and every worked example is fully visible, with a plain-language mode for non-engineers and a technical mode for engineers.
The five recurring decision axes
Nearly everything on this site reduces to one of these five judgment calls.
Must hold 100%, or just a preference?
Does this rule have to hold every single time, or is it a soft preference the model can weigh against other things? Guarantees need code (hooks, permission rules, orchestration, tool schemas); preferences can live in prompts or CLAUDE.md.
Trivial task, or complex and interdependent?
Is the task small, local, and unambiguous — or large, breaking, and spread across many files? Match the amount of process (plan mode, test suites, phased review) to the actual complexity and risk.
Shared or personal? Always-on or on-demand?
Every piece of configuration or guidance has a natural home: project file vs. personal file, always-loaded CLAUDE.md vs. on-demand slash command/skill, directory-wide vs. one-off @-reference.
Keep hard-won understanding, or start clean?
Context is a limited, degrading resource. Sometimes you want to keep everything the agent has learned (resume, fork); sometimes the understanding has decayed and a fresh, seeded context is worth more than continuity.
Change the tool/schema/error, or add more instructions?
When something goes wrong repeatedly (wrong tool picked, fabricated parameter, malformed output), the durable fix is usually structural — a better tool description, a schema change, a clearer error — not another round of emphatic wording.
Six teaching sections
Read these first — every worked example illustrates one of these.
Module 0
What This Exam Actually Tests
Judgment, not trivia — pick the mechanism whose strength matches what the situation needs.
Module 1
Claude Code Configuration
Match the mechanism to the requirement: instructions influence, hooks guarantee, and every kind of knowledge has a natural home.
Module 2
Working with the Agent
Match ceremony to complexity, iterate with concrete targets, and manage context as the limited resource it is.
Module 3
MCP and Tool Design
Agents pick tools from descriptions, fail well when errors are structured, and fabricate less when a tool's inputs are clearly sourced.
Module 4
Agentic Loops and the SDK
The loop runs on stop_reason, the model has no memory of its own, and hard guarantees live in the code wrapped around it.
Module 5
Structured Data Extraction
Enforce structure with a schema, stop fabrication at its source, model the real world honestly, and review by risk.
How to use this site as a reference
Concepts
One page per module. Every idea has a plain-language explanation and a technical explanation — toggle between them with the switch in the header, which applies site-wide. Jargon is underlined; hover or tap it for an inline definition.
Cheat sheet
A searchable, filterable "if you see this → do that" table. Good for a fast lookup once you already know the concepts.
Worked examples
58 unique reference scenarios (two duplicate pairs from the source guide were merged), each fully expanded: the situation, the recommended approach, why it works, and the weaker approaches people often try instead — all visible at once, filterable by module or decision axis.
Glossary
Every defined term in one place, plain and technical definitions side by side, linked back to where it's used.