Anthropic published the weightings for the Developer Foundations exam, and they’re strange. Not strange as in wrong. Strange as in nobody seems to have read them properly before writing a study plan.
Here they are:
| Domain | Share |
|---|---|
| Applications and Integration | 33.1% |
| Model Selection and Optimization | 16.8% |
| Agents and Workflows | 14.7% |
| Prompt and Context Engineering | 11.0% |
| Tools and MCPs | 10.6% |
| Security and Safety | 8.1% |
| Claude Code | 3.1% |
| Eval, Testing and Debugging | 2.6% |
One domain is a third of the paper. The bottom two together are 5.7%, which on a 53-question exam is about three questions.
Now go look at how the material is usually organised. Eight chapters, eight lessons, eight quizzes, roughly even. A week on Claude Code, a week on Applications and Integration. That’s a week each on three questions and on eighteen.
Why this happens
Partly it’s just how people write courses. Chapters feel like they should be similar sizes.
But there’s a second reason, and it’s more interesting: Claude is famous for prompting. If you asked a hundred engineers what “working with Claude” means, most would say writing good prompts. So Prompt and Context Engineering feels like the centre of gravity, and it’s tempting to spend real time there.
It’s 11%.
Meanwhile Applications and Integration is the least glamorous name on the list. Statelessness, stop_reason branches, prompt caching placement, config management, whether the batch API applies. It reads like plumbing. It’s a third of your score.
What the weighting is actually telling you
I think the weighting is Anthropic saying something fairly blunt about what a Claude developer does all day. Not prompt tuning. Integration work. Deciding whether the thing goes through the batch API. Noticing that a cache checkpoint is sitting behind a timestamp. Handling the response where stop_reason came back max_tokens and the answer is half a sentence.
That’s the job. The exam is weighted like the job.
There’s a secondary signal in the low numbers too. Claude Code at 3.1% doesn’t mean Claude Code is unimportant — it means the exam can’t test it deeply in a proctored multiple-choice format. Same with Eval and Debugging at 2.6%. These are things you demonstrate by doing, not by picking option C. Low weight is a statement about the format, not the skill.
What I’d actually do
Take the weights literally. If you’re putting in twenty hours:
- Seven hours on Applications and Integration
- Three and a half on Model Selection
- Three on Agents and Workflows
- Two each on Prompt/Context and Tools/MCPs
- Under two on Security
- An hour, total, on Claude Code and Eval combined
That last line makes people uncomfortable. An hour on two entire domains. But three questions is three questions, and the hour you’d spend getting from 60% to 90% on Claude Code is worth about one mark. The same hour spent on caching behaviour inside Applications and Integration is worth several.
Don’t skip them entirely — an hour is not zero, and the material is small enough that an hour genuinely covers it. Read the component list, understand the CLAUDE.md hierarchy, know what headless mode is for. Then stop.
The one caveat
Domain weights are percentages of the whole paper, not guarantees about any single sitting. A 53-question exam sampling at 2.6% might give you one Eval question or it might give you two. If you’ve written off a domain completely and get unlucky, that’s a couple of marks you didn’t need to lose.
So: proportional, not binary. Spend an hour where the guide says 3%. Don’t spend a week, and don’t spend nothing.
Weights are from Anthropic’s Claude Certified Developer – Foundations Exam Guide v1.0, effective July 2026. Anthropic revises guides without notice, so check the current one before you sit.