Hours studied and courses finished matter less than whether you have internalised six ideas. They recur across all five domains and all six scenarios.
1. When code beats a prompt
Prompts are guidance; code is enforcement. "Verify identity before refunding" works most of the time — and in production, the exceptions are unacceptable. Use prompts for preferences and formatting; use hooks, gates and validation for rules that cannot fail. "Enhance the system prompt to…" is almost always wrong when money, compliance or a hard business rule is involved. Domain 1.
2. How isolated subagents really are
A subagent starts from a blank page — not "limited context", none. Everything it needs must be written into its prompt every time. Multi-agent design sits in the largest domain (27%).
3. Tool descriptions that route correctly
Claude chooses tools by reading descriptions, not names. A good description says what the tool returns, what inputs it takes, when to use it versus a neighbour, and the edge cases. And remember: read-only data is an MCP Resource, not a Tool. Domain 4.
4. Where things get lost in long context
Information at the start and end of a long context is used more reliably than information in the middle. Fix it by restructuring — facts blocks, key findings first, scratchpad files — not by buying a larger window. Domain 5.
5. Which API, and when
| Option | Cost | Latency | Use when |
|---|---|---|---|
| Real-time API | Standard | Immediate | A user is actively waiting |
| Prompt caching | Up to 90% off cached input | Immediate | The same large context is reused |
| Message Batches | 50% off | Up to 24 hours | Background work nobody is waiting on |
The trap: "switch everything to Batch to save 50%". Anything blocking — pre-merge checks, live support — cannot wait a day. Repeated context plus a real-time need means prompt caching, never Batch.
6. Spotting the wrong answer first
The meta-skill. Distractors are written to tempt people who understand a concept at the surface. Learn their shapes:
- "Enhance the system prompt to…" when reliability is at stake.
- "Use the Batch API…" for a workflow where someone is waiting.
- "Give a single agent all the tools" — almost certainly wrong.
- "Have Claude report its confidence and escalate when low" — self-reported confidence is not calibrated.
- "Run Claude Code in the pipeline without extra flags" — missing
-p. - "Return 'Operation failed' on an MCP error" — no structured context.