CCAO-F
CCAO-F quick reference
The facts worth having in your head when you sit down. Print it, or read it on the way there.
Seven domains, weighted. Output Evaluation (21%) and Workflow Integration (16%) carry the most marks; Governance (15%) and Prompting (14%) follow. Facts that change often (prices, limits, model versions) are not tested as numbers; know the principle and that support.claude.com and the Anthropic docs hold the current figures.
Domain 1: Prompting and Task Execution
| Block | Ask yourself |
|---|---|
| Task | Named deliverable with an operational verb? |
| Context | Facts only I know written down? |
| Audience | Who reads it and what do they know? |
| Format | Length, structure, number of options? |
| Constraints | Must-haves and must-nots stated? |
- Role prompts help when they change reasoning, not just voice.
- Examples (few-shot) beat adjectives for tone, format and classification; keep them few and consistent.
- Iterate by follow-up (small fix), edit-and-resend (missing block), fresh chat (drift or changed goal).
- Feedback names what to keep and what to change; "try again" gives nothing.
- Task families and their failure modes: writing (generic tone), analysis (overconfidence), summary (smoothed disagreement), data (unverified figures).
Domain 2: Output Evaluation and Validation
- Hallucination = fluent, confident, fabricated fact. Not a tone or format problem.
- High-risk content: citations, precise numbers, quotations, niche or recent facts, your own organisation.
- Warning signs: unprompted specificity, tidy evidence, no hedging, internal inconsistency.
- Mitigations: supply facts, "only use provided material", ask for uncertainty flags, placeholders, source tagging.
| Tier | Output | Action |
|---|---|---|
| Glance | Notes, brainstorms | Read for sense |
| Scan | Internal emails, summaries | Check names, numbers, dates, promises |
| Check | Customer, management, figures | Verify each claim against a source |
| Audit | Legal, financial, regulatory, public | Primary sources, second reviewer, record |
- "Are you sure?" is not verification. Open the source. Prefer primary sources.
- Grounding: chat attachments are per chat; Project knowledge persists and may use retrieval; web search and Research return citations (origin, not truth).
- Rubric: three to five testable criteria, accuracy always included; put it in the prompt; ask for a self-check (reliable for structure, not for facts).
- Bias checks: swap a variable, ask the counter-case, who is centred, audit your premise. Register is set by example > named reference > concrete rules > vague adjectives.
- Numbers: prose = estimate; code run = computed. Most errors are specification errors (column, filter, unit, date, definition). Routine: anchor, recompute one, internal consistency, magnitude, definition.
Domain 3: Product and Model Selection
| Surface | Distinctive for |
|---|---|
| Web app | Full feature set, reference experience |
| Desktop | Shortcut, capture, local connectors, agentic file work (Cowork) |
| Mobile | Voice, camera, share sheet |
| Claude in Chrome | Acts on web pages; per-site permissions; confirmations |
| Claude Code | Software development only |
- Opus: complex, high-stakes. Sonnet: balanced default. Haiku: fast, high-volume. Version numbers change; tiers persist.
- Extended thinking: hard problems, slower; does not verify facts.
- Context window: all text in view, in tokens; long chats and big files approach it.
- Knowledge cut-off: recent facts need search or supplied material.
- Plans: Free, Pro, Max, Team, Enterprise. Usage limits are rolling and consumed by work done, not messages.
- Team/Enterprise = commercial terms (no training by default, admin control). Enterprise adds SSO, SCIM, audit logs, compliance API.
- Route by: kind of work, what Claude must see, difficulty and stakes, is the data allowed.
Domain 4: Workflow Integration and Solution Design
- Project = knowledge + instructions + conversations. Small knowledge fully in view; large knowledge retrieved (ask Claude to search). Shared Projects need an owner.
- Artifact = standalone output with versions; publish = public link (never confidential); interactive Artifacts run on the viewer's account. Generated files for distribution.
- Connectors use MCP; your access rights only; read is low risk, write needs confirmation; admins control availability; prompt injection = instructions hidden in retrieved content.
- Workflow levels: assist, draft-and-review, execute with checkpoints. Earn each level with evidence.
- Design for failure: placeholders, explicit uncertainty, structured output, self-check, reversible actions, sampling.
- Sharing: conversation link and published Artifact are public; generated files follow existing controls; shared Projects are organisation-scoped.
- Missing feature = ask the admin, never a personal account for company data.
Domain 5: Configuration and Knowledge Management
| Layer | Holds |
|---|---|
| Personal preferences | True of you everywhere |
| Project instructions | True of one body of work |
| Style | Voice and formatting, per conversation |
| Skill | How to do one kind of task, anywhere |
| Conversation | Latest message wins |
- Contradictions between layers cause inconsistency; remove them rather than rely on precedence.
- Instructions cover role, audience, format, tone, rules, rubric, uncertainty handling, how to use knowledge. Short.
- Memory: viewable, editable, Project-scoped, switchable; incognito = no history, no Memory. Distinct from Project knowledge and past-chat search.
- Knowledge curation: one current source per topic, dated names, headings, glossary, context note, owner, review cadence, test questions.
- Skill: SKILL.md + templates; progressive disclosure means the description decides when it loads; admins can enable org-wide.
Domain 6: Governance, Risk, and Responsible Use
- Consumer plans: training-use choice affects retention. Business plans: commercial terms, no training by default, admin-controlled retention.
- Minimise and anonymise. Never paste credentials. Public links are public. Connectors expose what you can see.
- Usage Policy: universal prohibitions (harm, deception, abuse, privacy), high-risk requirements (legal, health, finance, insurance, employment, housing need qualified human review and disclosure), platform rules (age, no circumvention, no account sharing).
- Fabricated reviews, fake personas, undisclosed endorsements = deceptive = prohibited.
- Your organisation's policy stacks on Anthropic's; stricter wins; escalate conflicts.
- IP: you own outputs under the terms; commercial plans add copyright indemnity; copyright protection depends on human authorship; respect licences on inputs; disclose where trust depends on it.
- Accountability never transfers. Decision rights: inform, recommend, act with checkpoint, autonomous, never. Meaningful review requires engagement, surfaced uncertainty, small batches, sampling. Keep an audit trail.
Domain 7: Troubleshooting and Optimisation
| Symptom | Likely cause | Fix |
|---|---|---|
| Different question answered | Vague task | Operational verb, concrete deliverable |
| Generic | Missing context | Facts, attachments, Project |
| Invented facts | Precision demanded without data | Supply facts, grounded only, placeholders |
| Wrong tone | Register unspecified | Example, reference, Style |
| Ignored instructions | Over-constraint, buried instruction | Prioritise, shorten, first and last |
| Old rule followed | Drift | Restate or fresh chat |
| Missed Project content | Retrieval | Ask to search, fix names and structure |
| Shallow reasoning | Light model, no thinking | Opus, extended thinking, reason first |
- Refusals: add honest context and purpose; never circumvent; repeated refusals mean reconsider the task.
- Limits: rolling allowance consumed by work; fresh chats and lighter models conserve it; multiple accounts breach terms; capacity messages are not your limit.
- Long work: chapters with handover summaries, Projects, lean attachments, outputs saved outside the chat.
- Patterns: clarify first, reverse brief, outline then draft, section by section, template fill, reason first, perspectives panel, counter-case, rubric and self-check, critique then revise, options menu, grounded only, source tagging, placeholders, explain for a reader, show the voice.
Ten things to remember walking in
- Specificity beats sophistication; name the missing block.
- Hallucination is fabricated fact, found most in citations, numbers, quotes and your own organisation.
- Verification depth is proportionate; open the source.
- Projects for recurring material, attachments for one-offs, Memory for context about you.
- Public links are public; confidential content travels as files through approved channels.
- Connectors act with your rights, confirm before writing, and can carry injected instructions.
- Business plans mean commercial terms; personal accounts are not a workaround.
- High-risk domains need a qualified human and disclosure; accountability never transfers.
- Opus for hard and high-stakes, Sonnet by default, Haiku for volume; look up the numbers.
- Diagnose the symptom, apply one fix, and after two failures rewrite the brief.