Free, reusable reference
Claude architecture skills matrix.
Use this matrix to teach, assess, or practise the decisions behind reliable Claude systems. Each domain connects one architecture question to observable evidence, a common weak signal, a focused guide, and original scenario practice.
Last source review: August 15, 2026 · Independent ClaudeQuest educational material · Not affiliated with or endorsed by Anthropic.
How to use the matrix
- 1. Name the decisionChoose the domain that changes the system outcome. Avoid scoring broad confidence without a concrete architecture choice.
- 2. Ask for evidenceLook for an explanation, design artifact, or fresh scenario result that another reviewer can inspect.
- 3. Re-test transferUse a different setting or combined constraint. Familiar wording measures recall; fresh decisions provide stronger evidence.
Seven decision domains
The review basis identifies the primary Anthropic source area used when ClaudeQuest last checked each domain. Product behaviour and certification details can change; verify current official information before making production or exam decisions.
- Core decision
- How should the goal, context, constraints, examples, and success criteria be expressed so the output can be evaluated?
- Observable evidence
- The learner can turn an ambiguous request into a testable instruction and explain why each constraint is present.
- Weak signal
- Longer prompts are treated as automatically better, while success criteria and failure handling remain implicit.
- Core decision
- Which facts, instructions, history, and retrieved evidence should enter the next model call—and what should be summarised or discarded?
- Observable evidence
- The learner can defend a context strategy using relevance, provenance, freshness, cost, and recovery requirements.
- Weak signal
- Every prior turn is retained by default, even when it hides instructions, increases cost, or preserves stale state.
Domain 3 of 7
Tools and MCP
Review basis: Anthropic tool-use guidance and Model Context Protocol concepts
- Core decision
- What is the narrowest tool contract that lets Claude make progress while keeping inputs, permissions, side effects, and failures observable?
- Observable evidence
- The learner can define validated inputs, bounded authority, idempotency or recovery, and a clear owner for every side effect.
- Weak signal
- A successful API call is treated as a reliable workflow without considering retries, partial completion, or duplicate actions.
- Core decision
- Where are the trust boundaries, and which actions need validation, least privilege, human approval, or a safe pause path?
- Observable evidence
- The learner can identify untrusted input, limit blast radius, and place approval before consequential or irreversible actions.
- Weak signal
- A refusal instruction is used as the only safety control while the system retains broad permissions.
- Core decision
- Which model, context, caching, batching, and output strategy meets the required quality, latency, and cost envelope?
- Observable evidence
- The learner can reduce repeated work without weakening a user-critical outcome and can name what should be measured.
- Weak signal
- Token count is optimised in isolation while retries, latency, evaluation, and human review costs are ignored.
- Core decision
- Is a deterministic workflow sufficient, or does the task need bounded agent judgment, delegation, shared state, and explicit stop conditions?
- Observable evidence
- The learner can choose the simplest viable architecture and explain state ownership, coordination cost, evaluation, and recovery.
- Weak signal
- More agents are added before task ownership, completion criteria, and failure handling are clear.
- Core decision
- How should repository guidance, permissions, planning, implementation, tests, and review gates be scoped for safe delivery?
- Observable evidence
- The learner can produce a bounded change, verify it with project evidence, and leave a reviewable handoff.
- Weak signal
- Broad repository authority is granted before the task, affected files, validation commands, and approval boundaries are understood.
Turn the matrix into your personal gap map.
Use fresh scenarios to find the first decision pattern worth repairing, then build a focused Claude learning route around the evidence.
Start the free diagnostic →