Claude Certified Architect – Professional

Not this time. Anthropic's "Building effective agents" names this case: orchestrator-workers suits tasks where you can't predict the subtasks, such as coding changes across a number of files that depend on the task. The orchestrator decides what has to change and hands out the pieces. A chain needs its steps known in advance. Running one prompt over every file wastes calls on the files that don't use the library and can't adapt to files that need something different. Routing chooses between fixed paths, but nobody knows the paths here.

Not this time. This is the routing workflow in Anthropic's "Building effective agents": classify the input and send it to a path built for it, and one of its examples is sending easy, common questions to a smaller model and hard ones to a more capable model, to optimise cost and speed. Then test the classifier, because a hard question sent the cheap way is the risk. The smallest model alone gives up quality where the client said not to. Shorter answers save little and hurt the hard cases. Stored answers go stale and fit a question only when it really is the same one.

Not this time. Anthropic's documentation says Claude's ability to pick the right tool degrades once there are more than 30 to 50 tools, and that the tool search tool fixes both problems: deferred tools stay out of the context until Claude searches the catalogue and loads the few it needs. A larger window holds the definitions but does nothing for picking among 300. Shorter descriptions make picking worse, because descriptions are what Claude chooses by. 14 agents hand the routing problem to the user. This is the progressive discovery against monolithic context trade-off the exam names.

Not this time. Embeddings capture meaning, and Anthropic's contextual retrieval article notes they can miss exact matches: for a query like an error code, an embedding search may return text about error codes in general and not the one asked about. BM25 matches the exact string, so combining both finds the E-4471 section and keeps the plain-language results. The same article reports that adding context to each chunk before embedding and indexing, plus reranking, cuts failed retrievals further. A bigger embedding model still matches by meaning, larger chunks blur what each one is about, and asking users to change how they search moves your problem onto them.

Not this time. Anthropic's research on the statistics of evals says a score means little without its standard error: on 100 questions at around 80%, the standard error is about 4 points, so a 95% interval is roughly 8 points either way and a 3-point gap can be noise. Because both prompts answer the same questions, a paired-differences analysis removes the variation in question difficulty and gives a tighter answer for free; if it is still unclear, add questions. Re-running until a high score appears is picking the result you want. A single overall judgment by a model adds another source of noise and hides which questions changed.

Not this time. Anthropic's guide on building evaluations ranks grading methods: code-based where it works, LLM-based for judgments that need nuance, and human grading as the slow, costly option to avoid where you can. For LLM grading it advises detailed rubrics, a scale or a fixed set of verdicts, letting the grader reason first, and testing that it is reliable before scaling it up, which is what checking against a human-graded sample does. Exact match can't grade tone or a correct answer phrased differently. 50 human grades a week is too few to stand for 5,000. A one-line question with no rubric gives judgments you can't trust or compare.

Not this time. The rule depends on a value in the tool's input, so it belongs in code that sees that input before the tool runs. The Agent SDK lets a hook deny or pass a call, and its canUseTool callback decides at run time, which is where you hold a large refund for approval and let small ones through. A rule in a prompt is followed most of the time, and that's not enough for money. Removing the tool loses the automation the client wants for small refunds. A weekly log review finds a wrong refund after it is paid. This is the human-in-the-loop validation the exam asks about: a person where the risk is, and none where it isn't.

Not this time. Anthropic's Bedrock page describes the global endpoint as dynamic routing across all available regions for maximum availability, which is the opposite of what the regulator asks. Regional endpoints serve data-residency needs: a single region you choose, or an inference profile that routes only within a geography such as the EU, and they cost 10% more than global ones. Removing names leaves the rest of the customer data travelling. A VPN changes where a request comes from, not where it is processed. Naming the premium is part of the answer: the client should hear the cost of the requirement.

Not this time. Anthropic's guide on success criteria asks for criteria that are specific, measurable and achievable: not "good performance" but, say, a score on a held-out set of real cases, set from a baseline or earlier tests rather than hope. For claims that means an accuracy figure on real claims, and an agreed path, such as sending uncertain cases to a person, for the rest. No model is right every time, so 100% is a promise the project would break. No number gives no way to tell when it's ready. A target chosen after launch means going live without one. Setting this expectation early is the stakeholder work the exam tests.

Not this time. Anthropic's model deprecations page says its dates apply to the platforms Anthropic operates, such as the Claude API, while partner-operated platforms, Amazon Bedrock among them, set their own retirement schedules, so the 2 products may have different deadlines. Requests to a retired model fail; nothing is moved for you. Anthropic gives at least 60 days' notice for public models and recommends testing the replacement on your own tasks well before the date. Planning around a single day for both, or after the date, risks one product failing on the day.

Not this time. Agent Skills package instructions and resources once and load them in stages: until a skill is triggered, only its name and description sit in the context, and the body loads when a task needs it, which Anthropic calls progressive disclosure. One skill gives every team the same copy and stops paying for the guide on requests that don't use it. A longer system prompt in each app keeps both problems. Fine-tuning is a heavy step for a document that keeps changing: every edit to the guide would mean training again. Claude can't read a wiki page it isn't given, so a line pointing at one does nothing.

Not this time. Claude Code reads settings from several levels and uses the highest one that sets a key. Managed settings, deployed through managed-settings.json, an MDM policy or the claude.ai admin console, sit at the top, and nothing in a developer's own files or a command-line --settings overrides them. That is the level for security policy. A shared project file and a user file are both below a developer's local settings or under their own control. CLAUDE.md is guidance Claude reads, not a permission rule that blocks a tool.