Claude Certification Guide

Changelog

Anthropic revises exam guides without notice, and study material goes stale quietly. This page records what changed and when — including corrections to material that was wrong here first. If something you studied has since been fixed, you'll find it below rather than nowhere.

  1. Content

    CCAR-F (Architect, Foundations) curriculum launched — 27 lessons, 70 questions

    The second certification track is live: Claude Certified Architect (Foundations), free and open at /ccar-f, with its own curriculum, diagnostic, practice bank, and mock exam.

    27 lessons across 5 domains — Agentic Architecture & Orchestration, Claude Code Configuration & Workflows, Prompt Engineering & Structured Output, Tool Design & MCP Integration, and Context Management & Reliability — plus a 70-question practice bank.

    On the domain weights: Agentic Architecture & Orchestration (27%), Claude Code Configuration & Workflows (20%), and Tool Design & MCP Integration (18%) are drawn from the published blueprint. Prompt Engineering & Structured Output (19%) and Context Management & Reliability (16%) are our best estimate of the remaining share — we don’t have those two figures confirmed against the current exam guide yet. If you have the official CCAR-F exam guide and these are off, that’s exactly the kind of thing this changelog exists to correct.

  2. Correction

    Corrected: MCP remote transport is Streamable HTTP, not sockets

    Lesson 8.2 and one practice question described the remote MCP transport as “network sockets.” That was wrong. The specification defines two standard transports: stdio for a locally launched subprocess, and Streamable HTTP for remote and multi-client access.

    Also worth knowing, since older material still repeats it: HTTP+SSE is deprecated, replaced by Streamable HTTP as of spec version 2025-03-26. If an answer option offers SSE as the current remote transport, that’s the dated option rather than the correct one. The practice question has been rewritten so SSE is the distractor.

    Source: MCP specification 2025-03-26

  3. Content

    Batch and caching figures replaced with verified numbers

    Several lessons said “a steep discount” where real, testable figures exist. Now stated:

    • Batch: 50% off both input and output tokens, applied automatically. 24-hour ceiling, not an estimate. Up to 100,000 requests or 256 MB. Results kept 29 days and returned out of order, matched by custom_id. No streaming inside a batch.
    • Caching: reads at 0.10× base input, writes at 1.25× for the 5-minute TTL or 2.0× for the 1-hour TTL. Reads reset the TTL.
    • The two stack — but the default 5-minute TTL expires partway through anything but a small batch, so use the 1-hour TTL inside one.

    The consequence worth carrying: a misplaced checkpoint is worse than no caching, because you pay the write premium on every call for a cache nothing ever reads.

    Source: Anthropic batch, pricing and prompt-caching documentation

  4. Correction

    Corrected: hook examples now use the real settings.json format

    Lessons 1.2 and 7.3 illustrated hooks with an invented Python decorator (@hook("pre_tool_use")). Hooks are configured as JSON in settings.json, events are PascalCase — PreToolUse, PostToolUse, UserPromptSubmit, Stop, SessionStart, PreCompact — and a matcher targets them at specific tool names.

    The handler receives the event as JSON on stdin and refuses either by exiting with code 2 or by printing hookSpecificOutput.permissionDecision. Exit 2 is the hard block: on events that can block it stops the call whether or not you print JSON.

    Both lessons now show the actual shape.

    Source: Claude Code hooks documentation

  5. Site

    Diagnostic added — ranks domains by cost, not accuracy

    Twelve questions, no feedback until the end. The report ranks domains by weight × (1 − accuracy), so scoring 0% on Eval and Debugging ranks below scoring 50% on Agents and Workflows. Bombing a 2.6% domain costs you almost nothing; being shaky on a 33.1% domain costs you the exam.

  6. Site

    Curriculum live — 25 lessons, one per sub-skill

    Every sub-skill in the blueprint now has its own lesson with objectives, worked examples, a practice scenario, a build exercise and its exam traps. Lesson length is sized by domain weight rather than split evenly.

  7. Blueprint

    CCDV-F Exam Guide v1.0 published, effective July 2026

    First published blueprint for Claude Certified Developer – Foundations. Eight domains, 25 individually weighted sub-skills, 53 questions, 120 minutes, 720/1000 to pass.

    The weighting is unusually skewed: Applications and Integration alone is 33.1%, the heaviest single domain on any of the four Claude certifications. Claude Code and Eval/Testing/Debugging together come to 5.7%.

    This is the version all material on this site is written against.

    Source: Anthropic exam guide v1.0