Getting certified — the Uniqus path
Anthropic Partner Certification · compiled from the official Partner Guide. Confirm fee, weightings and pass score on the Academy exam page before booking — Anthropic updates them.
Your checklist
- Create your account at anthropic-partners.skilljar.com — use your Uniqus email, not a personal one (sign-up is domain-gated)
- (Optional) Academy courses: Building with the Claude API · Claude Code in Action · Intro to MCP · Intro to Agents
- Register for the exam and pay $125
- Study: this app's domain pages + quiz + mock + revision sheet
- Pass the exam (≥ 720 / 1,000)
Exam structure & scoring
A ~301-level, scenario-based, proctored exam: 60 questions · 120 minutes · single session. Results in ~2 business days with a per-domain breakdown.
Domain weight → questions (of 60)
Passing score — scaled, not raw
Scored 100–1,000; pass = 720. Questions carry different weights, so there is no fixed number-correct. Passers report ≈43–44/60 raw as the practical bar — aim comfortably above it across all five domains.
Note: the Partner Guide says "multiple-choice", but verified passer reports include multiple-response items (each states how many to select) — train for both (the quiz does).
Recommended resources
- This app — your main study path: domain pages (01–06) with flip-card drills → adaptive quiz + timed mocks → 1-hour revision sheet
- Anthropic Claude Certified Architect — Full Practice Exams (Udemy) — the final readiness check: book the exam once you're consistently at 90%+
- Optional: Anthropic Academy core four — Building with the Claude API · Claude Code in Action · Introduction to MCP · Introduction to Agents
- Optional: Zero to Claude Certified Architect — Complete Exam Prep Series — free YouTube playlist
Exam overview — full study-kit notes optional deep reading
1. Exam logistics
| Fact | Value |
|---|---|
| Questions | 60 items — multiple-choice and multiple-response (each item states how many responses to select), scenario-based |
| Time | 120 minutes (~2 min/question) |
| Structure | 4 blocks × 15 questions, each block anchored to one production scenario drawn at random from a published bank of 6 |
| Scoring | Scaled 100–1,000, pass = 720 (≈ 69% raw; passers report 43/60 → 724, 44/60 → 738) |
| Fee | $125 USD per attempt |
| Delivery | Online proctored (Pearson) or test center — closed book: no Claude, no docs |
| Results | Not immediate — typically 2 business days to ~1 week |
| Validity | 12 months from award (Credly badge) |
| Retakes | Waits of 14 / 30 / 90 days after failures; max 4 attempts per rolling 12 months |
| Audience | Solution architects with ~6+ months building on Claude |
⚠️ Details have drifted since the March 2026 launch ($99→$125, 6→12-month validity, ProctorFree→Pearson, partner-only→public). Re-check the official Skilljar page before booking.
📊 Diagram — “Exam blueprint: bar chart of the five domain weights” (a version appears earlier on this page)
4. Officially OUT of scope — don't waste study time here
The exam guide explicitly excludes these topics. If a practice question tests them in depth, it's testing the wrong things:
- Fine-tuning Claude models or training custom models
- Constitutional AI, RLHF, or safety-training methodology
- Embedding models or vector-database implementation details (RAG internals)
- Computer use (browser/desktop automation)
- Vision / image inputs
- Streaming implementation details
- Rate limits and pricing specifics
- Cloud-provider configurations (AWS Bedrock, GCP Vertex, Azure)
- Prompt-caching implementation details — beyond knowing it exists (and its headline numbers)
Knowing the out-of-scope list is itself useful on exam day: distractor options sometimes invoke out-of-scope machinery (e.g. "fine-tune a model" or "add a vector database") — those are almost never the intended answer.
5. The exam mindset — how questions actually work
This exam is judgment-based, not recall-based. A typical question presents a production situation and four options that all technically work. You pick the one Anthropic considers most effective and reliable in production. As one passer put it: "every answer is workable, and the question turns on degree."
Four principles resolve most questions. Memorize them:
- Prompts guide. Systems enforce. A system prompt is probabilistic; a hook, schema, or code check is deterministic. When a behavior must happen (policy, compliance, safety), enforce it structurally — never rely on prompt instructions alone.
- Criteria explain the rule. Examples teach the boundary. State explicit criteria first; add few-shot examples to pin down the edge cases criteria can't capture.
- Don't just compress context — preserve decision-critical context. When summarizing or compacting, protect goals, constraints, key facts, assumptions, dates, sources, and prior decisions.
- Fix root causes before adding layers. Better tool descriptions before a routing classifier. Programmatic enforcement before a stronger prompt. Structural fixes before probabilistic patches.
And five recurring exam patterns reported by passers:
| Pattern | The tested insight |
|---|---|
| Prompt vs programmatic enforcement | Hooks and code enforce; system prompts guide |
| Silent failure | An empty tool result looks successful — tools must return explicit, machine-readable errors (isError: true) |
| Context isolation | Subagents start with no inherited context — everything they need must be passed explicitly |
| Batch API trade-offs | 50% cheaper, but only for latency-tolerant work (up to 24h window) |
| Tool scoping | Grant the minimum tool set required — least privilege for agents |
Timing strategy: ~2 minutes per question; commit within ~90 seconds, flag and return. Questions are revisitable within the exam.
Before you start — assumed background
The exam assumes working developer knowledge. You do not need ML theory — fine-tuning, RLHF, embeddings internals, and vision are officially out of scope. You do need these ideas cold:
| Term | What it means here |
|---|---|
| Token / context window | Models read and write tokens (~¾ of a word). The context window is the hard budget for everything: system prompt, history, tool results, output. |
| System prompt | Standing instructions with the highest authority — but still probabilistic guidance, not enforcement. |
| Messages API | Stateless HTTP: you resend the full messages history on every call. There is no server-side session. |
| Tool use | Claude returns a structured request (tool_use); your code executes it and returns a tool_result. Claude never runs anything itself. |
| Agent / agentic loop | Call model → execute requested tools → append results → call again, until stop_reason: end_turn. |
| MCP | Model Context Protocol — the open standard for connecting external systems (tools, resources, prompts) to Claude. |
| Claude Code | Anthropic's agentic CLI — configured via CLAUDE.md memory files, skills, hooks, slash commands, MCP servers. |
How to use this app
- This page — registration logistics + exam format (~30 min).
- Domain pages 01–05 — the content, ordered for learning (heaviest-weight and weakest-spot domains first). Flip the flashcards at the end of each page.
- Page 06 — the six scenarios as worked case studies; four appear on your exam form.
- Quiz — adaptive drills over a 338-question bank; missed concepts re-appear until you convert them twice.
- Quiz → Timed mock — 60 questions, 120-minute countdown, blueprint-weighted, no feedback until the report. Clear ~50/60 comfortably here, then finish with the Udemy mocks — book when you're consistently scoring 90%+.
- Revision — the 1-hour cheat sheet for the final days.
Flash-drill the exam facts (9 cards)
Tap a card to flip it — the same concepts the quiz drills adaptively.