Exam Facts
- Confirmed (Anthropic's own official Exam Guide PDF, v1.0, July 2026): 60 standalone items, 120 min, MC/MR, scaled score 100–1000, pass 720, $99, Pearson VUE (online proctored or test center), 12-month validity, free on-time renewal.
- Closed book: no Claude, no docs, no browser tabs, no notes, no phone, no second monitor.
- Retakes: wait 14 / 30 / 90 days after 1st/2nd/3rd fail; max 4 attempts / 12 months.
- Guide says "subject to change without notice" — re-verify before booking.
The 7 Domains
| Domain | Wt |
| 1 Prompting & Task Execution | 14% |
| 2 Output Evaluation & Validation | 21% |
| 3 Product & Model Selection | 12% |
| 4 Workflow Integration & Solution Design | 16% |
| 5 Configuration & Knowledge Mgmt | 12% |
| 6 Governance, Risk & Responsible Use | 15% |
| 7 Troubleshooting & Optimization | 10% |
Non-engineer exam: no code, no API, no SDK. That scope = Claude Developer/Architect.
What an LLM Is / Isn't
- Claude predicts fluent next text from a prompt — it is not a search engine or a database of verified facts.
- Confident tone ≠ accurate. Fluency and accuracy are separate things.
- No knowledge of "today," internal company data, or events after training cutoff unless you provide it or a connected tool (search) supplies it.
Tokens & Context Window
- Token ≈ chunk of a word. Context window = everything visible in one exchange (prompt + files + history + reply) — finite.
- Long chat drifting/inconsistent → summarize + start fresh, or move stable info into a Project. Don't just pile on more instructions.
Model Tiers
| Tier | Use |
| Haiku | Fast/cheap, high-volume, simple |
| Sonnet | Balanced, everyday default |
| Opus | Complex reasoning, high stakes |
"Always use the top model" = wrong-answer pattern. Match tier to task.
Seven Prompting Moves
- Be clear & direct — state task/audience/format/length
- Give it a role
- Show examples
- Let it reason step by step for real reasoning tasks
- Structure long/multi-part prompts (headings, sections)
- Give the material, not a description of it
- Say what "done" looks like (length, confidence flags, style)
Before / After + Decomposition
Before: "Summarize this report." → generic, wrong focus.
After: "5 bullets for a VP audience, focus on budget impact and timeline risk, skip methodology."
Before: "Fix our onboarding." After: decompose — list steps & drop-off points → propose alternatives → draft the emails.
- Multi-part request → split into ordered steps, verify each before building on it.
- Iterate the prompt, not just the output — name specifically what's wrong (tone, missing data, wrong audience). Keep a prompt journal.
Hallucination Checklist
- High-risk: specific numbers, dates, citations, quotes, very recent events.
- Ask: is it checkable? did you check it? does it match a doc you already have?
- Self-rated confidence from Claude ≠ validation. Verify against an authoritative source instead.
Exam pattern: confident cited claim → correct move is always "verify before sharing," never "send it, it sounded sure."
Escalate to Human / Engineer
| Situation | Escalate to |
| Legal/medical/financial judgment | Human expert |
| API integration, agentic system | Developer |
| Enterprise architecture decisions | Architect |
| Business workflow, built-in features | Stays with you (Associate) |
Projects vs. Chats
- Project = recurring task, shared instructions + knowledge sources + consistent team voice.
- One-off chat = single question, no reuse.
- Team getting inconsistent tone/format → shared Project, not longer individual prompts.
Artifacts
- Use for output you'll keep, edit, export, or share standalone (document, table, simple view).
- Inline chat text is fine for a quick, disposable answer.
- Structured data (table) > prose when the next step is comparing/sorting.
File Upload vs. Connector
- Fixed/reference doc that doesn't change often → upload.
- Live, changing source (shared drive, mailbox) → connector, if org policy allows.
- Connector to a live mailbox/drive can expose more than intended — check what's actually in it first.
Web Search — Helps vs. Hurts
- Helps: current events, prices, "status right now."
- Hurts/irrelevant: pure analysis of material you already gave it; highly sensitive internal-only questions.
Skills & Memory
- Skill = repeatable, saved way of doing a task, often shareable across a team.
- Memory = carries relevant detail forward across chats; convenience layer, not a substitute for Project instructions the whole team can see.
What Never to Paste In
- Secrets/credentials (passwords, API keys)
- Regulated personal data (health, financial, government ID) unredacted
- Other people's personal data without a clear need/right
- Confidential business info beyond what the task needs
- Anything your org policy specifically restricts
Exam pattern: spreadsheet with names/account numbers → anonymize first, not "upload, it's internal" or "tell Claude not to retain it."
Usage Policy & Ethics
- Anthropic's Usage Policy sets outer bounds; your org's policy is often stricter — both apply.
- Bias: outputs can reflect training-data bias — extra scrutiny for hiring/performance/lending decisions about people.
- Transparency: don't hide that a document started as an AI draft when it matters to the audience.
- Human stays accountable for consequential decisions — Claude assists judgment, doesn't replace it.
Verification Habits
- Match scrutiny to stakes: low-stakes draft = skim; customer/legal/regulatory = verify every claim + normal sign-off process still applies.
- Escalating to a human is a correct answer, not a failure.
- Edit/adapt output for its actual audience before sending.
Team / Admin Concepts
- Shared Project > individual reinvented prompts — one instruction set, one voice.
- Org-level admin sets access + data-source policy defaults — a second line of defense beyond individual judgment.
- Adoption: start with one scoped Project, write guardrails once, revisit configuration periodically (it goes stale).
Workflow Recipes
| Recipe | Prompt shape |
| Summarize | source + audience + length + what to prioritize |
| Extract | source + exact fields + table format |
| Draft | goal + audience + tone/examples + iterate |
| Analyze | data + specific question + reasoning first |
| Translate | text + audience/register + human review for anything customer-facing |
Troubleshooting Table
| Symptom | Fix |
| Generic output | Add specifics (audience/format/length) |
| Confidently wrong | Verify independently |
| "Forgets" instruction | Summarize + restart, or use a Project |
| Inconsistent across team | Shared Project |
| Slow / costly | Right-size model tier, decompose task |
Keyword → Answer
- "confident but unverified" → hallucination risk, verify
- "recurring task, shared context" → Project
- "standalone doc to keep/share" → Artifact
- "speed/cost over depth" → faster/cheaper model
- "complex reasoning, high stakes" → most capable model
- "sensitive data, internal use" → still redact/check policy
- "beyond your role" → escalate to human or engineer
- "current events" → web search / research mode
Top Traps
Confidence ≠ accuracy · self-rated Claude confidence ≠ validation · "best model for everything" is wrong · unredacted sensitive data upload is wrong · code/API answer choices are the wrong exam · unsupervised high-stakes automation removes accountability
When two options both sound right, pick the one that keeps a human verifying/accountable, not the one that trusts Claude's tone.