Claude Certified Architect – Professional

Skills

• — Translate business problems into AI solutions
• — Design input → processing → output → feedback architectures
• — Select workflow, agentic or augmented-LLM patterns
• — Design multi-agent orchestration and decomposition
• — Align to value pillars: efficiency, transformation, productivity, cost, SLAs

In practice

• A claims-triage request framed as 'use AI' becomes a workflow with one classification step and one summarisation step, plus a feedback loop from adjuster corrections — not an autonomous agent — because the SLA and audit requirements dominate.
• Choosing an agentic pattern for a research assistant and a strict workflow for invoice posting is the same judgement applied to different risk profiles.

Skills

• — Select models on measured trade-offs
• — Design system prompts, templates and guardrails
• — Apply zero-shot, few-shot and chain-of-thought where each fits
• — Optimise context windows and token use
• — Reuse prompts via caching, modular prompts and Skills

In practice

• A modular system prompt with a cached shared core and small per-tenant sections cuts cost and makes tenant-specific changes reviewable.
• Replacing a 'think step by step' instruction with extended thinking changes cost and latency; the architect decides where that is worth it.

Skills

• — Audit tool and agent configurations for capability bloat
• — Analyse authn/authz requirements and identify gaps
• — Trade off accuracy and latency with justified configuration
• — Select monitoring strategies for scale
• — Design RAG chunking, indexing and retrieval
• — Choose integration mechanisms: MCP, API/CLI, agent-to-agent
• — Compare progressive discovery with monolithic context

In practice

• An agent with write access to every internal system 'for flexibility' is a capability-bloat and authorisation finding; scoping tools per task and adding approval for writes is the fix.
• Chunking contracts by clause with metadata indexing beats fixed 500-token windows when queries are clause-specific; the opposite can hold for narrative documents.

Skills

• — Define accuracy, latency, cost, safety and security metrics
• — Build evaluation datasets and mixed-method test frameworks
• — Run A/B tests and iterate
• — Diagnose prompt failure vs hallucination vs model mismatch
• — Optimise token, latency and cost trade-offs
• — Monitor with logging and observability tools

In practice

• A golden set of 300 labelled examples, refreshed from production corrections each month, is what turns 'it seems better' into a release decision.
• A rise in refusals after a model upgrade is diagnosed by replaying the eval set against both versions, not by editing the prompt until complaints stop.

Skills

• — Implement guardrails and safety controls
• — Identify LLM failure modes and limitations
• — Apply human-in-the-loop validation
• — Design for GDPR, HIPAA and FedRAMP obligations
• — Address bias, fairness and transparency

In practice

• A patient-facing assistant in a HIPAA environment needs data-flow mapping, minimisation, access controls and a human review path before any model choice matters.
• Disclosing to customers that a response was AI-generated, and logging the human override rate, are governance controls you design in, not add later.

Skills

• — Run structured discovery and requirement gathering
• — Communicate architectural decisions and trade-offs
• — Manage feedback loops, expectations and SLAs
• — Document architectures and give implementation guidance
• — Support discovery, design, handoff, monitoring and iteration

In practice

• An architecture decision record that states the rejected options and why survives staff turnover; a slide that states the chosen option does not.
• Agreeing an SLA on p95 latency and a target human-override rate before build gives both sides something to measure the pilot against.

Skills

• — Configure Claude tools and environments for teams
• — Improve developer workflows with AI tooling
• — Support debugging and operational issue resolution

In practice

• Standardising CLAUDE.md, permitted tools and a review skill across twelve repositories is an enablement decision with security and consistency consequences.