Skip to content
Ford Prior Book a call

Writing · · 5 min read

I Took Anthropic's Easiest and Hardest Certs Back to Back (and Scored About the Same)

AI & agents Career

Originally published on LinkedIn.

Starting a few weeks ago, Anthropic now offers four Claude exams:

All four are pass/fail at 720 out of 1000 scaled, valid for twelve months, and Anthropic has signaled more tracks later in 2026 (so this map will probably be stale soon). *Topic of current debate!

Anyways, there's not a lot of non-Anthropic exam content available online, specifically about the Pro cert, so I'm going to offer my first impressions to my little community of builders.

First impression

I'm a big fan of AI competency frameworks. We need them because the tool is new and costs lots of money. I especially appreciate the leadership Anthropic and Claude has provided via the 4 D's, not to mention a library of 100% free high-quality courses known as the Anthropic Academy. For these reasons, I also like the idea of certification exams.

(An image appears here in the original article on LinkedIn.) The 4D framework in a nutshell. Source: https://www-cdn.anthropic.com/334975cdec18f744b4fa511dc8518bd8d119d29d.pdf

My preparation

  1. My professional background is 2 years of building agentic systems, and 10 yrs of software development before that. I've passed 3 AWS associate-level exams. I used to identify as a poor test taker. I studied 1 hour/day for 3 weeks.
  2. I couldn't find anything online, so I used roughly $200 of Claude credits to pull in all published exam guide and practice questions (which I analyzed deeply for patterns using Fable) and expanded on the guides using the Claude Cookbook and Developer Docs to make sure I covered everything. That gave me a corpus of content.
  3. Then, I layered in "exam hacks" based on established patterns whereby keywords map to certain canonical answers, with expected red herrings, which I learned while studying for the AWS exams. I talked to Claude about it, leading to a deep research report on "AWS test taking psychometrics and construct-irrelevant variance", which I then asked Fable 5 to map carefully onto the Claude exam corpus.
  4. Lastly, I asked Claude to divide the corpus into modules, and for each module to create 1) flashcards, 2) infographic prompts (which I passed to ChatGPT's Images 2.0 model), and 3) "architect cheat sheets" consisting mostly of decision-trees.
  5. Bundle it up, add login/security, bundle it up, and ship to web (learnclaude.app) and iOS (see on App Store).

Here's the end-to-end depiction (F5 stands for Fable 5; Opus 4.8/Sonnet 5 for everything else):

(An image appears here in the original article on LinkedIn.) Two notable exceptions to using Claude/AI for this use case: 1) I very much enjoyed creating my own strange acronyms and mnemonic devices, such as "Every Eskimo Buys It Under Kover" to remember the principles for writing MCP tool descriptions, and 2) I did NOT trust Claude with generating practice exam questions.

Associate Foundations Cert - how AI fluent are you?

IMO, the Associate - Foundations (CCAO-F) cert is built for a frontline knowledge worker driving Claude to tasks like drafting, reviewing, and iterating on artifacts such as business plans or the product proposals. My first impression was that the exam tested only 3 things:

  1. Applying the 4D framework across 3 interaction modes: Automation (finely-scoped tasks), Augmentation (human collaboration), Agency (goal-oriented tasks).
  2. Your ability to call BS on factual inaccuracies on 100-level topics such as model features and context. Specifically, they want you to know that 1) Opus is different from Haiku on cost, latency, and reasoning depth, and 2) context should be groomed, massaged, fine-tuned, and refreshed constantly. These questions came as checkboxes "Which feedback is most appropriate if a coworker says x about AI? [Select all that apply]".
  3. Whether you're disciplined about product development or tempted by "AI shortcuts", i.e. have you forgotten how to do old-school requirements gathering & build in proper mechanisms? These came in the form of awkward sequencing scenarios that "Here are A, B, C, D, E steps in a proposed Claude workflow. Put them in order".

(An image appears here in the original article on LinkedIn.) General format of workflow sequencing challenges.

Pro Architect Cert - are you ready to build agentic systems?

Where the Associate cert was about usage, the Pro Architect cert is about building the systems themselves. For the most part, it came down to 3 things:

  1. Deep recall of capability matrices given a scenario with a half dozen variables in play, how well can you surface technical details (see app on web or ios), at approximately the same depth and breadth as the AWS Associate-tier certs (i.e. not necessarily route tables and CIDR math, but long capability matrices for various use cases, such as when to use the seven stop_reason values).
  2. Prioritization under constraint. For example, the exam had these ~6-element boolean checklists ("how important is each optimization for MVP? [High/Low]") that gauged how well you could make lots of small decision among competing options.
  3. Spotting key "operators" for quick architectural triage, which allows you to apply the above matrices in the field. In the format of exam questions, these appear as little keywords or phrases buried in a scenario that deterministically orient the whole question toward one answer type. For example, if the exam explains a high-performing production AI workflow & asks for the first measure to optimize on cost, the answer will be related to prompt caching.

(An image appears here in the original article on LinkedIn.) Mention of "cost optimization" + "first remedy" for high-performing workflows = cache.

There's a whole family of these. Another that comes to mind is: if it mentions cost optimization for latency-tolerant workflows, the answer will usually be Batch API and its fifty-percent discount, not a faster mode. Another is scenarios where "guaranteed" or "deterministic" behavior is needed, the answer is never CLAUDE.md instructions and instead a PreToolUse hook. And so on.

TL;DR

Overall, unsurprisingly, Anthropic isn't just testing memorization of the docs; it's testing whether you can think like someone who builds with Claude responsibly in the wild, when ethics and cost and risk are all in play. Within that frame, the Associate Foundations cert's marquee question is about recall and order (use a process!) and the Pro Architect's cert is about prioritization and judgement (play chess, not checkers). That's my first impression, at least.

P.S. About my 100% free app (Learn Claude)

If you're serious about learning the content for whatever reason, check out my iOS app (or the web version at learnclaude.app. The basic info:

  1. It's called Learn Claude and is a 21-day, self-paced guide to getting fluent enough to pass the Architect-level exams within 1 calendar month.
  2. The recommended flow: each day Monday-Friday you read a short lesson in the app, then do a hands-on exercise on your own machine with a Claude session beside you. The app is "the map"; the free companion repo is the territory. Also included in your backpack are spaced-review flashcards (including a terse deck with mnemonics, reference docs on tools, hooks, MCP, and subagents.
  3. Your progress syncs across devices and it works offline. Free, iPhone-only (for now). Web version if your an Android person: learnclaude.app. All the companion exercises are free & open source at github.com/ford-at-home/claude-code-companion.

Your honest App Store review genuinely me make the app better! 🙏 Link to app.