Domain 3: Claude Code configuration and workflows

If you use Claude Code daily, this is the domain that feels safe, because it is the one part of the exam about a tool you already use. That familiarity is the risk: habits that work fine on your own machine are not always the ones the exam marks as right.

For example, when two CLAUDE.md files disagree, it is natural to assume the more specific one wins, the way a CSS cascade would. It doesn't. Every file that applies is loaded and joined together, and the model reads the lot, so neither is guaranteed to win on a conflict. If a rule has to hold, it belongs somewhere the model cannot reinterpret.

That one fact bears on much of Claude Code Configuration & Workflows (20% of the CCAR-F, about 12 questions). The rest hangs on a single distinction.

Loaded, invoked, or run without you

Every piece of Claude Code configuration falls on one side of a line. Some is loaded into every session whether you ask for it or not. Some waits to be called, or for a matching file to be touched. And some of the domain is about running with nobody at the keyboard at all. Sorted that way, the modules look like this:

• Loaded for you: the CLAUDE.md hierarchy (3.1), where guidance lives and how several files combine, and path-specific rules (3.3), conventions that apply only to some kinds of file.
• Called when needed: commands and skills (3.2), the reusable workflows you invoke by name.
• How you work: plan mode (3.4), and iterative refinement (3.5), getting output back on track when it drifts.
• Nobody at the keyboard: non-interactive mode and the CLI flags (3.6a), and CI workflows (3.6b).

Modules 3.1 to 3.3 are taught back to back, because none of them is clear on its own. Take a convention that applies to one kind of file across a whole codebase. The two obvious places to put it are both wrong, for different reasons, and working out why takes all three modules.

Some of what the hierarchy module covers is mundane and still tested: which files are committed and which stay on your machine, how splitting a long CLAUDE.md across imported files tidies it up without saving any context, and what comes back after /compact and what doesn't. What comes back after /compact explains an instruction that held for an hour and then stopped.

Commands and skills look like the same thing from the outside, since both give you a slash command. Underneath, one is a single prompt file and the other is a directory that Claude can also pick up on its own when it matches what you are doing. Drop a loose file in the wrong folder and you get nothing, with no error to tell you so.

Two judgement calls

Plan mode is tested on how many reasonable ways there are to do a task. Difficulty doesn't enter into it. A nasty bug with an obvious location and an obvious fix doesn't need a plan, however long it takes. A change that sounds small but could sensibly land in several different places does, because picking the wrong one is the expensive mistake. Large, repetitive changes are a third case the module handles separately.

Refinement questions put one wrong answer on offer whenever Claude reads an instruction differently from run to run.

The module covers the techniques that do work, and whether several corrections should go in one message or one at a time. That depends on whether the fixes affect each other, and the answer sticks once you have been asked about it a few times.

In a pipeline

The last two modules move Claude Code into CI. In an interactive session a missing answer is a pause. In a pipeline there is no one to answer, so the job needs to be told up front how to run, what shape to return, when to stop, and what it may do without asking. The flags for all of that are tested by name, -p included (it means print, not plan, whatever it sounds like), and a couple of the options you'll see look plausible and don't exist.

The pipeline half asks how the run that writes code and the run that reviews it should relate. A session that has just written something reviews it badly, for reasons Domain 4 explains in detail, so the exam keeps the two apart. It also asks what a review should already know when it starts, so it doesn't rediscover the same problems on every push, and the difference between adding to the system prompt and replacing it, which are two flags with different outcomes.

What the modules cover

The seven modules go into the scopes CLAUDE.md files load from and which are committed, what a skill holds on disk compared with a command, the glob that makes a rule apply across a whole tree, the flags one by one, and how a pipeline keeps its reviewer independent. With Full Access, Spaced teaches each in full and schedules the practice around your evenings. This domain is largely recognition at speed, and that comes from being asked again on days when you have half forgotten the material.