CCAO-F Revision Notes
Part of the CCAO-F track. Domains, weights and objective IDs come from registry.yaml;
the human-readable summary is exam-facts.md. Product facts are as of
2026-10-24. Use these notes in the final week, after working through the
study guides.
The exam in five lines
- 60 items, multiple-choice and multiple-response, 120 minutes, Pearson VUE (online or test centre).
- Seven domains; Domain 2 (Output Evaluation and Validation) is the largest at 21%.
- Questions test judgement in realistic business situations: the best next action, the right feature, the right level
of review.
- No coding or API knowledge is needed.
- Updated guide due 2026-11-17; updated exam content from 2026-12-07.
Domain 1 · Prompting and Task Execution (14%)
- Strong prompt = goal + context + audience + constraints + format + success criteria.
- Say what to do (not only what to avoid) and why the rule exists.
- Examples show style and format faster than adjectives; vary them.
- Long material first, question last; label the parts.
- "If it is not in the material, say 'not stated'" reduces invented facts.
- Decompose multi-phase or high-stakes work; check between steps; ask Claude to plan first.
- Iterate diagnostically: name the problem, find the cause in the prompt, change one thing, compare. Edit and
regenerate when the first attempt went off track.
- Task types: analysis → evidence and reasoning; research → scope and sources; drafting → audience and voice;
brainstorming → volume and variety, then rank.
Domain 2 · Output Evaluation and Validation (21%)
- Accurate ≠ complete. Check both, against the brief.
- Hallucination warning signs: figures, dates, clause numbers, quotes, URLs or names you did not supply; claims
beyond the source; recent events with no source.
- Validation means an outside source: primary document, recalculation, independent cross-check, expert review.
Self-rated confidence and "are you sure?" are not validation.
- Citations must be opened and confirmed.
- Review level rises with: stakes, external audience, irreversibility, specialist expertise, novelty, decisions
about people.
- Bias checks: swap names or groups; look for coded language; ask whose perspective is missing.
- Audience adaptation: executives (decision first), staff (what changes, what to do), customers (plain, accurate
commitments), technical (detail, assumptions). Check for drift when shortening.
- Compare outputs against criteria set in advance.
- Formats: inline (quick), artifact (substantial, reusable, shareable), structured data (feeds another tool).
- You own what you send.
Domain 3 · Product and Model Selection (12%)
| Signal |
Feature |
| Recurring work, shared files, consistent team behaviour |
Project (instructions + knowledge) |
| Multi-source investigation with citations |
Research (paid plans, web search on) |
| Quick current fact |
Web search |
| Substantial deliverable or small tool to edit and share |
Artifact |
| Content lives in Drive, Gmail, Calendar and so on |
Connector |
| Same procedure every time, across workspaces |
Skill |
| Continuity of personal preferences |
Memory or chat search (where enabled) |
| Do not save this chat |
Incognito chat (still follow policy) |
- Haiku: fastest, cheapest, simple high-volume tasks. Sonnet: balanced default for everyday work. Opus: most
capable of the three, for complex reasoning. Fable 5.1: Anthropic's most capable widely released model, premium
tier (as of 2026-10-24).
- Choose the cheapest model that meets the quality bar, tested on real examples.
- Context window: everything Claude can consider in one conversation. Up to 1M tokens for the newest models on paid
plans; part is reserved for the reply. Long chats use more of your limits and can drift.
- Restart (drifted or new topic) · Summarise (long but healthy) · Persist (needed across chats).
- Research runs on Claude Opus 5.5 whatever model the chat uses (as of 2026-10-24).
Domain 4 · Workflow Integration and Solution Design (16%)
- Good first use case: frequent, time-consuming, text-heavy, checkable, low risk, permitted data.
- Claude can analyse interview notes into steps, pain points and contradictions; people resolve the contradictions.
- Augment → automate a step → redesign. Each step needs stronger evidence and controls.
- A named human checkpoint before anything external, irreversible or affecting individuals.
- Measure a baseline before the pilot. Track a quality measure as well as time.
- Prototype with Projects and artifacts; test on easy, typical and awkward examples.
- Communicate: problem and baseline, what Claude does and what people do, evidence including failures, limitations
and controls, cost, decision needed. No overselling.
- Escalate integrations, automation at scale, public-facing products and systems of record to Architects,
Developers and IT.
Domain 5 · Configuration and Knowledge Management (12%)
- Project = instructions (how to behave) + knowledge (what to know) + chats (the work).
- Projects on all plans (Free: up to five); RAG mode extends knowledge capacity on paid plans; Team and Enterprise
sharing with "Can view" or "Can edit".
- Chats in a Project do not share their conversations with each other; put shared facts in knowledge.
- Knowledge hygiene: current, authoritative, one version per document, clear names with dates, nothing
unnecessary or restricted.
- Upload stable reference material; connect live, changing content.
- Connectors: Team and Enterprise owners enable them and can set actions to allow, needs approval or blocked; each
user authenticates individually; custom (remote MCP) connectors are not verified by Anthropic.
- Instruction template: purpose, audience, knowledge use, format, rules (with reasons), when unsure.
- Instructions are not a security control.
- Maintain: owner, review date, change log, test set; retest after any change or model update.
Domain 6 · Governance, Risk, and Responsible Use (15%)
- Appropriateness test: permitted by policy? data allowed? who is harmed if wrong? human decides? comfortable
explaining it?
- Classify before you paste: public, internal, confidential, restricted.
- Minimise: redact, pseudonymise (key kept outside the tool), aggregate, synthetic data, excerpts. Usually better
than either skipping the task or uploading everything.
- "Don't remember this" is not a control. Neither is a personal account with training switched off.
- Commercial plans (Team, Enterprise, API): no training on customer data by default. Consumer plans (Free, Pro,
Max): training only if the user opts in. Incognito: not in history or memory; not used for training (as of
2026-10-24).
- Policy unclear → ask and pause. Mistake with data → report promptly.
- Ethics: accuracy, fairness, transparency, accountability, privacy, intellectual property, impact on people,
over-reliance.
- Decisions about people stay with qualified humans.
Domain 7 · Troubleshooting and Optimization (10%)
| Symptom |
First thing to check |
| Generic |
Missing context or audience |
| Wrong format |
Format unstated or conflicting |
| Invented facts |
No source, no fallback rule |
| Missed input |
Too much material, poor labelling |
| Ignores a rule |
Vague, buried or contradicted instruction |
| Out of date |
Old knowledge file, no web search |
| Inconsistent |
Duplicate files, conflicting instructions, different access |
| Worse over time in one chat |
Context drift: summarise and restart |
| Shallow on hard task |
Model too light, no room to reason |
- Reproduce → change one thing → retest on saved examples → log the change.
- Fix recurring issues in configuration, not chat by chat.
- Efficiency levers: Projects (cached reuse), Skills and templates, right-sized models, shorter chats, batching,
structured outputs, review matched to risk.
- If repeated fixes fail, reconsider task fit or escalate.
Last-day checklist