Pass the Claude Certified Architect Exam with Confidence

System prompt vs. hard gates

Using system prompt instructions to enforce business logic instead of programmed hard gates that can't be talked around.

Few-shot examples vs. tool descriptions

Adding few-shot examples when the root cause is an inadequate tool description — the model needs better context, not more examples.

Over-engineering before simplifying

Reaching for classifiers or ML pipelines before exhausting simpler, deterministic approaches that solve the same problem.

Sentiment escalation vs. policy gaps

Escalating based on sentiment signals rather than explicit policy gaps or stated customer preferences.

Too many tools per agent

Assigning more than 4–5 tools to a single agent — beyond this threshold, tool-selection reliability degrades silently.

Message Batches for blocking flows

Using the Message Batches API inside workflows that require real-time responses — batches have a 24-hour processing window.

Self-review by the same session

Applying self-review by the same session that generated code, rather than routing to an independent instance for genuine critique.

Text content vs. stop_reason

Checking response text content to determine loop termination instead of reading the stop_reason signal directly.