Check yourself answers

• Ques‐
• tion Answer
• 1.1 4 scenarios, from a bank of 6
• 1.2 Agentic Architecture & Orchestration, 27%, about 16 items

1.3 Any three of: fine-tuning, billing, MCP hosting, computer use, vision, streaming, pricing, OAuth, cloud configs, tokenization, prompt caching details

1.4 The total scaled score

2.1 It stops after the first text-plus-tool-call turn. Check stop_reason instead 2.2 Findings and source metadata not passed in the synthesis prompt; outputs not structured to separate content from metadata

2.3 Hook, at the pre-tool-call point, redirecting to escalation

2.4 Resume if most context is still valid, and name the three changed files; start fresh with a summary if much is stale

3.1 Rewrite both descriptions with purpose, inputs, examples and when to use each 3.2 isError, errorCategory, isRetryable and a human-readable message 3.3 tool_choice: {"type": "any"}

3.4 In your user-level ~/.claude.json instead of the project .mcp.json 4.1 .claude/skills/<name>/SKILL.md in the repo

5.4 Schemas check structure, not whether values are right or consistent 6.1 Escalate immediately with a structured hand-off

6.2 A persistent case facts block outside the summarised history 6.3 Report both with sources and dates, marked as contested

6.4 Accuracy by document type and field, calibrated thresholds, stratified sampling