What is the fundamental limitation of Claude that requires external code to manage multi-step tasks?	Claude is stateless; it does not remember previous calls, run loops, or execute code itself.
In the agentic architecture, what is the purpose of the "Inspect" step in the Call-Inspect-Execute-Append-Call loop?	To check the model's `stop_reason` (e.g., checking if it equals `tool use`) to determine if code needs to run next.
When should a coordinator use Sequential decomposition versus Adaptive decomposition?	Sequential decomposition is for steps that depend on the previous output; Adaptive decomposition is for open-ended investigation.
What is the primary function of the `.claude.md` file in Claude Code, and what is the difference between the project-level and user-level versions?	It provides context/instructions. Project-level applies to the repo; user-level applies globally to the user account.
What is the recommended, reliable method for forcing a model to output structured data, and why is it superior to asking it in the prompt?	Defining a structured tool/schema and using the API's function calling mechanism. This forces the model to adhere to the structure.
When should you use the `messages` API endpoint for large-scale, asynchronous processing, and what is the key benefit?	When the conversation history is very long. The benefit is that the model only processes the necessary context window, saving tokens and improving relevance.
What is the primary benefit of using the `messages` API endpoint for large-scale, asynchronous processing, and what is the key benefit?	When the conversation history is very long. The benefit is that the model only processes the necessary context window, saving tokens and improving relevance.
What is the purpose of the `messages` API endpoint for large-scale, asynchronous processing, and what is the key benefit?	When the conversation history is very long. The benefit is that the model only processes the necessary context window, saving tokens and improving relevance.
When defining multiple tools that perform similar functions (e.g., getting user info), what is the required fix to prevent schema defects?	Write tool descriptions like API documentation, including: 1) what the tool does, 2) when to use it over siblings, 3) an example invocation, and 4) a list of potential errors.
What is the required structure for a tool's error response to allow the model to write a clean recovery plan?	A JSON object containing `is_error` (Boolean), `category` (string), `retryable` (Boolean), and optionally `retry_after_ms` (integer).
According to the transport selection rule, when should you choose STDIO over SSE?	Choose STDIO whenever the server can live in the same process tree as the client (same machine).
To combat the "lost in the middle effect" in long context windows, how should durable facts (like IDs or amounts) be handled?	Identify the facts, copy them into a structured "case block," and re-anchor this block at the very end of the context on every turn.
When implementing prompt caching, what types of content are worth caching using a cache control breakpoint?	Identical, reusable sections like the system prompt or few-shot examples. (The user turn is not worth caching.)
What is the primary pattern for reliable agent escalation?	The agent must perform an explicit confidence check, handing off to a human if its stated confidence falls below a threshold or if it detects ambiguity.
What is the danger of using summarization when a conversation contains critical IDs or amounts?	Summarization compresses information with lossy heuristics and can silently drop durable facts.
In a multi-agent system, what is the correct retry strategy after a sub-agent fails?	Only retry the specific sub-agent that failed, not every sub-agent, as side effects may have already occurred.
Where must team-shared configuration rules be placed to ensure they are version-controlled?	In the project file at the root of the repository, not in the user's local home directory (`user-clawed.md`).
What is the "vague adjective trap" to avoid when refining prompts?	Do not use vague adjectives like "careful" or "thorough" to fix false positive problems, as the model uses numbered, categorical rules.