# Study Plans

Week-by-week plans for each certification with hour allocations derived from the official blueprint weights, hands-on builds and booking gates.

import { Tabs, TabItem } from '@prosefly/astro-components';

All plans use the same rules: hours follow blueprint weight; you build something real; you only book after a timed practice exam scores ≥ 80% with no domain below 65%.

<Tabs>
  <TabItem label="CCAO-F (4 weeks, ~25 h)">

| Week | Hours | Focus | Deliverable |
| --- | --- | --- | --- |
| 1 | 7 | D2 Output Evaluation (5.5 h) · D7 Troubleshooting (1.5 h) · Anthropic Academy: Claude 101, AI Capabilities and Limitations | Personal evaluation checklist; three deliberately-elicited hallucinations documented |
| 2 | 7 | D4 Workflow Integration (4 h) · D1 Prompting (3 h) | One real workflow mapped: where Claude adds value, where humans gate |
| 3 | 7 | D6 Governance (4 h) · D5 Configuration (3 h) | A configured Project with instructions + knowledge; a one-page acceptable-use summary |
| 4 | 4 | D3 Product/Model Selection (3 h) · Practice exam + review (1 h+) | Timed 60-item practice exam; book if ≥ 80% |

  </TabItem>
  <TabItem label="CCDV-F (6 weeks, ~35 h)">

| Week | Hours | Focus | Deliverable |
| --- | --- | --- | --- |
| 1 | 6 | D1 Applications & Integration part 1: Messages API, streaming, errors, retries · Academy: Building with the Claude API | Script calling the API with streaming and backoff |
| 2 | 6 | D1 part 2: caching, batches, vision/PDF, config management · D7 Claude Code (1 h) | Caching demo showing `cache_read_input_tokens`; a batch job |
| 3 | 6 | D2 Model Selection & Optimization · D8 Eval (1 h) | Cost model spreadsheet; 20-case golden set with exact-match + rubric |
| 4 | 6 | D3 Agents & Workflows · Academy: Intro to Subagents | Agentic loop dispatching on every `stop_reason`; one subagent |
| 5 | 6 | D4 Prompt & Context Engineering · D5 Tools & MCP · Academy: Intro to MCP | Custom tool + minimal MCP server; structured output with validation-retry |
| 6 | 5 | D6 Security & Safety · Practice exam + review | PreToolUse hook blocking destructive commands; timed 53-item practice exam; book if ≥ 80% |

  </TabItem>
  <TabItem label="CCAR-F (4 weeks, ~40 h)">

| Week | Hours | Focus | Deliverable |
| --- | --- | --- | --- |
| 1 | 11 | D1 Agentic Architecture (full) · Scenarios 1 & 3 | Coordinator/subagent research system with structured errors and `stop_reason` termination |
| 2 | 10 | D2 Claude Code Config (6 h) · D4 Tool Design & MCP (4 h) · Scenarios 2 & 4 | Repo with CLAUDE.md hierarchy, settings, hooks, a Skill, a subagent, `.mcp.json` |
| 3 | 10 | D3 Prompt Eng & Structured Output (6 h) · D5 Context & Reliability (4 h) · Scenarios 5 & 6 | Extraction pipeline with JSON schema, validation-retry, per-type metrics; headless CI review |
| 4 | 9 | Anti-patterns drill · Practice exam · weak-domain review | Timed 60-item practice exam; book if ≥ 80% |

  </TabItem>
  <TabItem label="CCAR-P (6 weeks, ~50 h)">

| Week | Hours | Focus | Deliverable |
| --- | --- | --- | --- |
| 1 | 8 | Re-validate CCAR-F baseline (skim its 5 domains, take its practice exam) | ≥ 85% on CCAR-F practice exam |
| 2 | 10 | D3 Integration & RAG (full) | RAG pipeline with defensible chunking, hybrid retrieval, reranker, recall@k and faithfulness evals |
| 3 | 8 | D1 Solution Design (5 h) · D7 Dev Productivity (3 h) | ADR set for a reference architecture; managed-policy settings for a team |
| 4 | 8 | D4 Evaluation, Testing & Optimization | Eval harness with A/B comparison and CI regression gate; observability dashboard spec |
| 5 | 8 | D5 Governance, Safety & Risk (4 h) · D2 Models/Prompting/Context (4 h) | Risk register + DPIA outline; caching architecture with stable-prefix layout |
| 6 | 8 | D6 Stakeholder Communication & Lifecycle · Practice exam | Discovery interview guide, decision matrix, lifecycle plan with exit criteria; timed 63-item practice exam; book if ≥ 80% |

  </TabItem>
</Tabs>

## Daily micro-routine (15 minutes)

1. Read five Lexique entries you cannot define from memory.
2. Re-answer two practice questions you got wrong yesterday, aloud, explaining each distractor.
3. Skim one decision table in the Decision Frameworks appendix.

## Booking gate checklist

- [ ] Timed practice exam ≥ 80% overall.
- [ ] No domain below 65%.
- [ ] Can explain all ten anti-patterns and their correct alternative (technical exams).
- [ ] Can recite the model lineup, prices and Fable 5.1 breaking changes (technical exams).
- [ ] Can list the mandatory human-review triggers (all exams).
- [ ] Downloaded the current official Exam Guide and confirmed weights have not changed.
- [ ] Partner-domain email confirmed; Pearson VUE account created; OnVUE system test passed.
- [ ] Reminder set for renewal at month 10.

## Final week – day-by-day

The last seven days are for consolidation, not new material. Do not cram; sleep matters more than one more topic.

| Day | Focus | Output |
| --- | --- | --- |
| −7 | Full timed practice exam (exam 1), cold | Baseline score + per-domain breakdown |
| −6 | Attack the two weakest domains from the breakdown | Re-do every missed item aloud, explaining each distractor |
| −5 | Second weakest domains + all anti-patterns / model facts drill | Recite the 10 anti-patterns and their fix; the model table |
| −4 | Practice exam 2 (harder), timed | Compare per-domain to day −7; note any domain `<65%` |
| −3 | Decision frameworks + Lexique gaps | Every decision table explainable from the signal word |
| −2 | Light review: skim your own notes; re-answer 10 previously-missed items | Confidence pass, no new topics |
| −1 | Rest. Logistics only (OnVUE system test, ID, environment) | Night-before checklist done |
| Day 0 | Exam | Arrive/settle 30 min early |

:::tip[Exam signal]
If practice exam 2 (harder, multi-constraint) is within 5 points of exam 1, you are ready. A large drop means you pattern-matched exam 1's wording rather than the underlying principle — go back to the signal-word index.
:::

## 40 must-know facts (flashcards)

Cover the answer; recall it cold. These are the highest-yield facts across all four exams.

1. Pass = 720/1000 scaled; 120 minutes; no guessing penalty.
2. Retakes: 14 / 30 / 90 days; max 4 per rolling year.
3. Validity 12 months; free on-time renewal; Credly badge.
4. Model IDs: `claude-fable-5-1`, `claude-opus-5`, `claude-sonnet-5`, `claude-haiku-4-5`.
5. Prices per MTok in/out: Fable $10/$50 · Opus $5/$25 · Sonnet $2/$10 · Haiku $1/$5.
6. Context: 1M for Fable/Opus/Sonnet; 200k for Haiku. Max output 128k / 64k (Haiku).
7. Fable 5.1: forced `tool_choice` → 400. Use `auto`+instruction, `strict`, or `output_config.format`.
8. Fable 5.1: append-only history; thinking blocks bound to producing model; 30-day retention (no ZDR); not Priority Tier.
9. Only Haiku 4.5 still uses `budget_tokens`; others use `{"type":"adaptive"}`.
10. `effort`: `low | medium | high | xhigh`; not on Haiku 4.5.
11. Cache: read 0.1×, write 1.25× (5-min) / 2× (1-hour); min 1024 tokens (2048 Haiku); break-even ≈ 2 reads.
12. Batch API: 50% off, results ≤24 h; for overnight/bulk.
13. `stop_reason`: `end_turn | tool_use | max_tokens | stop_sequence | pause_turn | refusal`.
14. Loop terminates on `stop_reason`, never on parsing prose or an iteration cap.
15. `tool_result` goes in a **user** message, matched by `tool_use_id`.
16. Tool errors: structured `is_error` (category, retryable) — never empty success.
17. Retry only 429/5xx/529 with backoff + jitter + `retry-after`; fix 4xx.
18. Rate-limit dimensions: RPM, ITPM, OTPM — whichever binds first.
19. Structured output: `output_config.format` + `strict: true`, then validate + retry.
20. Critical business rule → deterministic **hook** (PreToolUse, exit 2 blocks), not a prompt sentence.
21. Escalate on explicit request or beyond-capability — never sentiment or self-reported confidence.
22. Report **per-segment** metrics; aggregate accuracy hides failing segments.
23. Self-review in a **fresh session / different model**, never same-session.
24. 4–8 tools ideal; beyond ~10 use tool search + `defer_loading`.
25. Remove an unneeded destructive tool (least privilege) — not log/confirm it.
26. MCP primitives: Tools (model), Resources (app), Prompts (user); + sampling, roots.
27. MCP transports: stdio (local) vs Streamable HTTP (remote, OAuth 2.1 + PKCE).
28. MCP is model↔tool, not agent↔agent.
29. MCP connector = server-side, no local harness; pass user-scoped token.
30. Workflow = predefined code paths; Agent = model directs its own steps. Prefer the simplest.
31. Six workflow patterns: chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, augmented LLM.
32. CLAUDE.md precedence: managed → user → project → subdirectory; `CLAUDE.local.md` personal; `@` imports.
33. settings.json scopes: managed > flags > local > project > user.
34. Skill (procedure) vs CLAUDE.md (conventions) vs command (one-liner) vs subagent (isolated task) vs MCP (external system).
35. Plan mode for multi-file/architectural/irreversible; direct for single obvious edits.
36. Headless: `claude -p --output-format json|stream-json`, `--allowedTools`, `--permission-mode`.
37. RAG for large/changing corpus + citations; long context for small stable; fine-tuning for style not facts.
38. Confident-but-wrong after a doc refresh → suspect retrieval/indexing first.
39. Hybrid retrieval (dense + BM25) + reranker; query rewriting/HyDE for vague queries.
40. Compliance: GDPR (EU personal data, DPIA), HIPAA (PHI + BAA), FedRAMP High (Bedrock/Vertex), SOC 2, ZDR.

## Night-before checklist

- [ ] OnVUE system test passed on the *same* machine and network you will use.
- [ ] Government photo ID ready; name matches the Pearson VUE registration exactly.
- [ ] Quiet, private room; desk clear of notes, phones, second monitors.
- [ ] Webcam, microphone and stable internet confirmed; power adapter plugged in.
- [ ] Browser extensions/VPN that could block the proctor disabled.
- [ ] Exam start time and time zone double-checked; plan to launch 30 minutes early.
- [ ] Downloaded the current Exam Guide once more to confirm no weight changes.
- [ ] Sleep. No new studying after the −1 review.

## Retake strategy by domain score

If you fail (or want to convert a marginal pass into a strong renewal baseline), use the score report's per-domain breakdown to target the retake. Respect the 14-day minimum before the first retake.

| Weakest-domain score | Interpretation | Retake plan |
| --- | --- | --- |
| 0–40% | Conceptual gap, not test nerves | Re-study that domain page fully + its practice questions; rebuild the hands-on deliverable; wait the full interval |
| 41–60% | Knows the topic, misses the trade-off | Drill the decision tables and signal-word index for that domain; re-do missed items explaining every distractor |
| 61–79% | Careless / time pressure | Timed drills on that domain only; practise eliminating distractor patterns fast |
| ≥80% but overall failed | Broad thin coverage, no single weak domain | Full mixed-domain practice under time; focus on multi-constraint "FIRST/MOST/TWO" items |

:::caution[Retake anti-pattern]
Do not simply re-sit sooner hoping for easier questions. The item bank rotates; a failed attempt on the same weak domain wastes one of your four yearly attempts. Fix the domain, verify with a timed practice exam, then rebook.
:::