CCAR-F Study Plan — Day by Day

Three variants. The 3-week plan is the default; the 2-week variant compresses it, the 4-week variant adds a consolidation week.

Assumed effort: 90–120 min on weekdays, 2–3 h on weekend days. If you have less time, cut the "stretch" line first — never cut the question banks or the mocks, because those are what move the score.

The daily learning loop

1. READ the chapter once, fast, skipping code blocks.        (~25% of the time)
2. CLOSE it. Reproduce the "Decision reflexes" box on paper. (~20%)
3. Run the matching flashcard deck cold.                     (~25%)
4. Do the domain qbank items assigned for the day.           (~30%)
5. LOG every miss: what I picked | what was right | rule violated.

Re-reading with no recall step feels productive and is nearly worthless. Steps 2–4 produce the score.

Where your hours should go

Time budget should track the blueprint, not your comfort:

Domain 5 gets slightly under-weighted deliberately: its content overlaps heavily with Domains 1 and 2 (error propagation, escalation, subagent context), so you absorb much of it while studying those. Domain 1 gets over-weighted because it is the largest block and the hardest to intuit.

3-Week Plan (default)

Week 1 — Agentic architecture and tools (Domains 1 + 2, 45% of the exam)

End-of-week check: you can explain, without notes, why an agent misroutes between analyze_content and analyze_document, and what you'd change first.

Week 2 — Claude Code and prompting (Domains 3 + 4, 40% of the exam)

End-of-week check: given a config requirement, you can name the correct file and its exact path without hesitating.

Week 3 — Output, reliability, and consolidation (Domains 4 + 5, then everything)

2-Week Compression

Same sequence, doubled up. Drop all stretch items and Labs 3–4.

⚠️ On the 2-week plan, do not skip Ch. 13. Scenario familiarity is the cheapest points on this exam.

4-Week Extension

Run the 3-week plan for days 1–21, then add a consolidation week:

Scoring grid — track this after every mock

The exam reports percent-correct by domain. Reproduce that:

On the scaled score: the pass mark is 720 on a 100–1,000 scale, and Anthropic does not publish the raw-to-scaled mapping — it comes from a standard-setting study, so no percentage is guaranteed to pass. Treat 44/60 (73%) as the floor and 48/60 (80%) as a comfortable margin. Because the exam is criterion-referenced, your score depends only on your own performance; there is no curve to game and no benefit to being fast.

Exam-day tactics

Time. 120 minutes / 60 items = 2 minutes each. That's generous for this exam's style, so budget deliberately: first pass at ~90 seconds per item, flagging anything you don't resolve quickly. That leaves ~30 minutes for the flagged set plus a final sweep.

Read the scenario once, carefully, then stop re-reading it. All items in a block share the same scenario. Extract the constraints on first read — target metric, tool list, team size, what's already failing — and work from your own notes after that.

Read the last sentence of the item first. "What is the most effective first step?" and "Which two changes should the architect make?" call for different answers from identical stems. Multiple-response items always tell you how many to select — count them before you commit.

"Most effective first step" penalizes over-engineering. When one option is a small, targeted, root-cause fix and another is a redesign, the small fix wins — unless the stem says the small fix has already been tried.

Deterministic beats probabilistic when consequences are financial or compliance-related. A hook, gate, or programmatic check beats any prompt instruction, few-shot set, or "add it to the system prompt" option. Prompt instructions have a non-zero failure rate; that phrasing is the guide's own.

Watch for options describing things that don't exist. The sample questions use fabricated flags and files as distractors (CLAUDE_HEADLESS, --batch, a commands array in .claude/config.json). If you can't picture the exact file path or flag, be suspicious.

Distrust "more context" answers. A bigger context window does not fix attention dilution, and resuming a stale session does not fix stale tool results. Splitting work into focused passes does.

Distrust confidence and sentiment. Any option resting on the model's self-reported confidence for a routing decision, or on customer sentiment as a proxy for complexity, is a distractor. Confidence is only usable when calibrated on a labeled validation set.

Don't change an answer without a statable reason. "It feels wrong now" is not a reason. "I misread which tool the coordinator owns" is.

Blame the layer the evidence implicates. If logs show the subagents did exactly what they were asked and the output is still incomplete, the fault is in the coordinator's decomposition, not the subagents.

The ten Core Reflexes

These are in Chapter 01 §4 with full explanations. Memorize the list; you should be able to name which one applies to any given item.

1. Deterministic enforcement > prompt instruction, whenever an error costs money or compliance.
2. Fix the root cause with the lowest-effort, highest-leverage change first.
3. Tool descriptions are the primary mechanism for tool selection — fix them before restructuring.
4. Least privilege on tool distribution; scope the common case, route the exception.
5. Subagents have isolated context and inherit nothing — pass what they need explicitly.
6. Structured errors with category and retryability > generic failure strings; never suppress, never kill the whole workflow.
7. Attention dilution is fixed by splitting passes, not by enlarging context.
8. Schemas eliminate syntax errors, not semantic ones — validate meaning separately.
9. Explicit, testable criteria + 2–4 targeted examples > adjectives like "be conservative."
10. Self-reported confidence and sentiment are uncalibrated; don't route on them.