CCAR-P Exam Prep: Claude Certified Architect Professional

A team taught an invoice to file itself with an autonomous agent when a deterministic workflow would have done the job, and it cost them. This lecture covers the pattern ladder (augmented LLM, workflow, agent), the rule the exam keeps testing (use the simplest pattern that meets the requirement), decomposition, and two worked questions.

A five-agent research assistant turned out to have two agents doing the thinking and three passing notes between them. This lecture covers over-orchestration, the honest case for splitting (context isolation, parallel breadth), the business value pillars, and two worked questions on when a split earns its keep.

A retail team shipped a support assistant that looked flawless in the demo, then the bill came in for re-sending the same eight thousand tokens on every call. This lecture covers model selection as a capability, speed and cost trade-off, prompt caching mechanics, context budget discipline, prompt reuse with Skills, and three worked questions.

A support agent had four tools wired into it, including a refund tool and a delete tool that nobody remembered using, both live to any stranger who typed into it. This lecture covers capability bloat as an attack surface, auth gaps a security review will find, choosing between MCP, a plain API and agent-to-agent, and three worked questions from the largest domain on the exam.

A company refreshed its docs on a Friday and its support assistant answered questions all weekend, fast, confident and wrong. This lecture covers chunking trade-offs, indexing and re-indexing discipline, retrieval strategy matched to data and query shape, contextual retrieval, and when RAG is the wrong tool.

A retrieval agent fleet got worse for a whole week and not a single alarm went off, because nobody had ever logged retrieval quality or per-stage latency. This lecture covers the layers you actually watch in an LLM system, what changes at scale, the accuracy-latency trade-off, and defending a configuration in SLA language.

A tighter prompt rewrite read cleaner and shipped on a Friday, then quietly dropped a legally required disclaimer on a slice of replies. This lecture covers golden sets, metric selection including safety and security, grading in layers, and building an eval set that carries its own failure and adversarial cases.

We swapped in a cheaper, faster model because it looked fine in the demo, and it scored lower than the version it replaced once we ran it against the golden set. This lecture covers running an honest A/B test, a diagnosis tree for locating the real cause of a regression, the correct order of optimization moves, and regression protection.

A payments team ran a refund agent that issued small credits on its own and stopped for a human above a limit. One week the agent misread a currency field and queued a refund with far too many zeros on it, and the person on the gate caught it in seconds. This lecture covers deterministic against probabilistic controls, a failure-mode inventory for LLM systems, guardrail layering, and where a human belongs, decided by risk tier rather than by comfort.

A finished healthcare project could not launch, because nobody at the table could say where the patient data goes the moment a nurse types it in. This lecture covers the data-flow map you draw before naming a regulation, GDPR, HIPAA and FedRAMP requirements at architect altitude, mapping them to real controls, and ethical AI as an architecture decision.

A correct design died in one meeting because it was pitched as a diagram instead of a decision. This lecture covers discovery questions that surface the real constraint, the trade-off narrative, setting an SLA for a probabilistic system without promising determinism, and the handoff document set.

One widened permission setting, checked into a shared config, gave forty engineers the same open door into production. This lecture covers where team configuration belongs, why guidance is not enforcement, the workflow patterns worth standardizing, and how much study time this smallest domain actually deserves.

Four different Claude systems failed, and every postmortem opened with someone blaming the model. In all four cases the model was fine. This lecture takes apart four real failure shapes (context rot, stale retrieval, tool-scope creep, an eval-blind regression) and leaves you with a three-question diagnosis tree you can run on anything that breaks.

An engineer who knew the material lost a question because the deciding clause was buried three sentences in. This lecture dissects the three published sample questions in full, works nine original questions across the domains the samples skip, and teaches the reading discipline that actually wins multiple-response items.

Sixty-three questions, a hundred and twenty minutes, and every one of them a decision rather than a definition. This lecture covers pacing, the two-pass strategy, multiple-response discipline, reading a score report, the retake ladder, and deciding you're ready with no official practice exam to check yourself against.