Claude Code mind maps
Everything the CCA-F exam asks about Claude Code, on one page. It starts with which of the six exam scenarios bring Claude Code questions. Then six mind maps, each in the form "what you see, where the fault is, the fix", with the rule to remember, the exam-guide line it comes from, and one practice question.
Domain 3 · 20%Guide 3.1–3.6, plus 1.7, 2.5, 4.1, 4.5, 4.6, 5.4About 30 minutes
Which of the six scenarios ask about Claude Code?
The exam shows you 4 scenarios out of a bank of 6. Three of the six are about Claude Code:
| # | Scenario | Domains | Claude Code questions? | Maps to study |
|---|---|---|---|---|
| 2 | Code Generation with Claude Code | D3, D5 | Yes, the core one. CLAUDE.md, slash commands, rules, plan mode | 1, 2, 3, 5 |
| 5 | Claude Code for Continuous Integration | D3, D4 | Yes. -p, JSON output, reviews, false positives | 4 (and 1) |
| 4 | Developer Productivity with Claude | D2, D3, D1 | Partly. Built-in tools, exploring a codebase, MCP servers | 6, 2, 5 |
| 1 | Customer Support Resolution Agent | D1, D2, D5 | No (Agent SDK) | — |
| 3 | Multi-Agent Research System | D1, D2, D5 | No (Agent SDK) | — |
| 6 | Structured Data Extraction | D4, D5 | No (API) | — |
Guide §5: “4 scenarios drawn from a bank of 6.” Scenario 2: “custom slash commands, CLAUDE.md, plan mode vs direct execution.” Scenario 5: “automated reviews, test generation, PR feedback, minimize false positives.” Scenario 4: “explore codebases, built-in tools (Read/Write/Bash/Grep/Glob), MCP servers.”
First, the picture: where things live
Most Claude Code questions are really one of two questions: who should get this? (shared or personal) and when should it load? (always, for some paths, or only when asked).
The six maps at a glance
| # | Question | Usual answer | Practise |
|---|---|---|---|
| 1 | Where does an instruction live? | Everyone, always: project CLAUDE.md. Some file types: .claude/rules/ with paths:. Only when asked: a command or skill | Where an instruction lives |
| 2 | Plan mode or just do it? | Many files or a design choice: plan mode. One clear fix: direct | Plan or just do it |
| 3 | Claude keeps getting it wrong. How do you iterate? | Show examples, write tests first, or let Claude interview you | Examples, tests first or interview |
| 4 | Claude Code in CI | -p, JSON output, an independent reviewer, explicit criteria | Claude Code in CI, Self-review, Batch or now |
| 5 | The session got long or stale | Resume, fork, start fresh with a summary, or compact | Resume, fork or fresh, Large codebase |
| 6 | Which built-in tool? | Names: Glob. Contents: Grep. Small change: Edit | Built-in tools |
A yellow row in each map is the runner-up: the case where a different answer wins because one fact in the question changes.
Mind map 1: Where does an instruction live?
Ask two questions: who needs it (the team or just me) and when (always, for certain files, or only when someone asks).
| What you see | Where the fault is | Fix |
|---|---|---|
| A new developer's Claude ignores the team's coding standards | The standards are in someone's ~/.claude/CLAUDE.md, which isn't shared | Move them to the project CLAUDE.md (root or .claude/CLAUDE.md) |
A convention applies only inside one folder, like src/api/ | It's in the root file, so it loads everywhere | A directory CLAUDE.md in that folder |
| Test files are spread across many folders and all follow one convention | A directory file can't follow them around | .claude/rules/ file with paths: ["**/*.test.tsx"] |
The root CLAUDE.md is huge and hard to maintain | Everything is in one file | Split it with @import or topic files in .claude/rules/ |
Everyone needs a /review command after cloning | It's in ~/.claude/commands/ | .claude/commands/ in the repo |
| A long workflow fills the main chat with noise | It runs in the main context | A skill with context: fork in its SKILL.md |
| Claude doesn't seem to follow an instruction at all | You don't know which files loaded | Run /memory to see the loaded files |
| It must happen every single time, with no exceptions (for example, run the formatter after each edit) | Any text instruction is followed most of the time, not always | A hook, not CLAUDE.md |
CLAUDE.md is always loaded, for standards everyone needs. Rules with paths: load only for matching files. Commands and skills load only when asked. Personal files are never shared. Use a hook when it must never be skipped.Guide 3.1: “User-level is not shared via version control. @import syntax for modular references. /memory command verifies which memory files are loaded.” Guide 3.2: “Skills = on-demand task workflows; CLAUDE.md = always-loaded universal standards.” Guide 3.3: “Glob-pattern rules beat directory CLAUDE.md for conventions spanning directories.”
- A section in the root
CLAUDE.md - A
CLAUDE.mdin each service's migrations folder - A
.claude/rules/file withpaths: ["**/migrations/**"] - A
/migrationskill developers run
Mind map 2: Plan mode or just do it?
| What you see | Where the fault is | Fix |
|---|---|---|
| The change touches dozens of files, or needs a design decision (like service boundaries) | Starting edits now risks rework | Plan mode: explore, then design, then change |
| There are several valid ways to do it | Claude would pick one without comparing | Plan mode |
| One clear fix in one place, like adding a single validation check | Planning would only add time | Direct execution |
| Exploring the codebase floods the main context with search output | Discovery runs in the main session | Use the Explore subagent. It returns a summary |
| The question says the work is large from the start, and an option says “start direct, switch to plan if it gets complex” | The complexity is already known | Plan mode now. “Switch later” is the trap |
Guide 3.4: “Plan mode: large-scale changes, multiple valid approaches, architectural decisions, multi-file mods. Direct execution: simple well-scoped changes (single validation check). Explore subagent isolates verbose discovery output, returns summaries.”
- Enter plan mode and design the fix first
- Make the change directly
- Send an Explore subagent to find related messages
- Ask Claude to interview the developer first
Mind map 3: Claude keeps getting it wrong. How do you iterate?
| What you see | Where the fault is | Fix |
|---|---|---|
| A prose description of a transformation keeps being misread | Words are ambiguous | Give 2 or 3 concrete input and output examples |
| You need proof the code behaves correctly | “Looks right” isn't checked | Write tests first, then share the failures with Claude |
| The task is in an unfamiliar area and the requirements aren't clear | Claude would guess the missing parts | The interview pattern: Claude asks questions before building |
| Several problems affect each other | Fixing them one by one undoes earlier fixes | Describe them all in one message |
| The problems are independent | One big message mixes them | Fix them one after another |
Guide 3.5: “Concrete input/output examples communicate transformations better than prose. Test-driven iteration (write tests first, share failures); the interview pattern (Claude asks questions first); single message for interacting problems, sequential for independent ones.”
- Rewrite the description in more detail
- Show three real log lines and the exact output wanted for each
- Use the interview pattern
- Switch to plan mode
Mind map 4: Claude Code in CI
This is scenario 5. It mixes Domain 3 (the CLI) with Domain 4 (review quality).
| What you see | Where the fault is | Fix |
|---|---|---|
| The CI job hangs waiting for input | Claude Code is running interactively | claude -p "…" (or --print) |
| Findings must become inline PR comments | Free text can't be parsed reliably | --output-format json with --json-schema |
| Reviews and generated tests ignore the project's standards | The CI run has no project context | Put testing standards, fixtures and review criteria in CLAUDE.md |
| The session that wrote the code also reviews it, and misses bugs | It shares the reasoning that made the bug | Use an independent review instance |
| Too many false positives, and developers stop trusting the review | Vague instructions like “be conservative” | Explicit categorical criteria (what to flag, what to skip), with examples |
| A 14-file PR gets uneven, contradictory feedback | One pass spreads attention too thin | Per-file passes, then one cross-file pass |
| An overnight tech-debt report costs too much | It runs in real time but nobody waits for it | Message Batches API (50% cheaper, up to 24 hours) |
| The check blocks a merge | Batch has no speed guarantee | Keep it real time, not batch |
-p and JSON output, give context through CLAUDE.md, review with a fresh instance, and cut false positives with explicit criteria. Use batch only when nobody is waiting.Guide 3.6: “-p / --print = non-interactive mode. --output-format json + --json-schema for machine-parseable findings. The session that generated code is less effective at reviewing its own changes than an independent instance.” Guide 4.1: “‘Be conservative’ / ‘only high-confidence’ don't improve precision like categorical criteria.”
- Add “only report high-confidence issues” to the prompt
- List which issue types to report (bugs, security) and which to skip (style), with an example of each
- Run the review three times and keep issues found twice
- Use a larger model
Mind map 5: The session got long or stale
| What you see | Where the fault is | Fix |
|---|---|---|
| You want to carry on yesterday's conversation | A new session starts empty | --resume <session-name> |
| You want to try two approaches from the same starting point | One session can only follow one path | fork_session |
| You resume after files were changed outside the session | Claude still believes the old file contents | Resume and tell Claude what changed |
| Most of the old tool results are now out of date | Resuming brings back stale results | Start a new session with a structured summary |
| The context is nearly full in a long session | Old turns take up space | /compact |
| Answers drift to “typical patterns” instead of what was found earlier | The context has degraded | Save findings to a scratchpad file; hand verbose digging to subagents |
| A long multi-step job crashes halfway and has to start again from zero | Progress lives only in the conversation | Keep a structured state file (a manifest) of what's done, and resume from it |
/compact for space, and a scratchpad file for findings that must survive.Guide 1.7: “Starting a new session with a structured summary is more reliable than resuming with stale tool results.” Guide 5.4: “Scratchpad files persist key findings across boundaries; subagent delegation isolates verbose output. Structured state persistence (manifests) for crash recovery; /compact to reduce context usage.”
--resumethe session and carry on- Start a new session with a short summary of yesterday's conclusions
fork_sessionfrom yesterday's session- Resume and run
/compact
Mind map 6: Which built-in tool?
This is mostly scenario 4 (developer productivity).
| What you see | Where the fault is | Fix |
|---|---|---|
Find files by name, like every *.test.ts | — | Glob (file path patterns) |
| Find where a function is called | — | Grep (searches file contents) |
| Change one line in a file | — | Edit (matches a unique piece of text) |
| Edit fails because the text appears more than once | The match isn't unique | Read the file, then Write it back |
| Claude reads every file before starting, and the context fills up | It reads everything up front | Grep for the entry points, then Read along the imports |
| Claude uses Grep when a better MCP tool exists | The MCP tool's description is weak | Improve the MCP tool's description (see MCP map 3) |
Guide 2.5: “Grep = content search; Glob = file path patterns; Read/Write = full file ops; Edit = targeted change via unique text match. When Edit fails on non-unique match → Read + Write fallback. Build understanding incrementally: Grep entry points → Read to follow imports/trace flows.”
- Glob for every source file and Read them all
- Grep for the payment entry point, then Read the files it imports, one by one
- Read the README and answer from it
- Ask the developer which files matter
Three traps that catch people
1 · Personal file or project file?
If the question says “every developer”, “on clone” or “new team member”, the answer is in the repo: project CLAUDE.md, .claude/commands/, .claude/rules/ or .mcp.json. Anything under ~/ is only for you.
2 · Directory CLAUDE.md or a rules file?
If the files sit in one folder, a directory CLAUDE.md works. If they're spread across folders but share a name pattern (tests, migrations), a .claude/rules/ file with paths: wins.
3 · Commands: the exam's answer vs today
Today's Claude Code merges commands into skills, but the guide still teaches .claude/commands/ for a shared slash command. On the exam, follow the guide. More on Exam vs 2026.
How to study this
Related: Domain 3 hub · Claude Code lab · MCP mind maps · Exam vs 2026 · all traps · CCA-F start here