Answer Patterns — Anthropic House Style
Remember
These patterns break most 50/50 ties. They come from the real question bank and from people who passed.
The 10 patterns
| # | Situation in the question | Anthropic's preferred answer |
|---|---|---|
| 1 | Customer asks for a human / is angry | Escalate to human. With a structured handoff summary |
| 2 | Output format is wrong or inconsistent | Add few-shot examples |
| 3 | Claude ignores or misuses a tool | Improve the tool description |
| 4 | Metadata (sources, dates) lost between agents | Structured output: content + metadata separated |
| 5 | Rule must ALWAYS hold (compliance, limits) | Enforce in code / hooks, never only in prompts |
| 6 | Context too big for one agent | Delegate to subagents, pass summaries back |
| 7 | Data quality broken at step 2 | Fix step 2, not a repair step at the end |
| 8 | Agent hallucinates a parameter value | Tool description: "must come from a prior tool call" |
| 9 | Big batch job, not urgent | Batch API (50% cheaper). Urgent → real-time API |
| 10 | Task is complex or risky | Plan / explore first, then implement |
Micro-decisions that repeat
- Grep = search file contents. Glob = find files by name/path.
- Edit keeps failing on a file → Read the file, then Write the whole file.
- Files changed on disk → tell the agent which files changed (targeted re-read), or start fresh if everything changed.
- Small rename in code → just tell the agent. Do not restart the session.
- Compare two approaches from a shared starting point → fork the session (one branch each).
- Subagent has too many tools → give each subagent only its 4–5 relevant tools.
- Delegating to a subagent → put the needed findings in its prompt. Subagents do not share memory.
- Micromanaging a subagent fails → give goals and quality criteria, not step-by-step orders.
Words that signal the answer
| Word in question | Likely answer theme |
|---|---|
| "guarantee", "must never", "compliance" | Hook / code enforcement |
| "inconsistent format" | Few-shot examples |
| "loses track of sources" | Structured claim–source mapping |
| "wrong tool chosen" | Better tool descriptions |
| "context window", "50K tokens", "slow after 30 turns" | Summarize old turns, keep recent verbatim |
| "overnight", "not urgent", "10,000 documents" | Batch API |
| "irreversible", "production", "destructive" | Human confirmation with full details shown |
Next: Exam Day Plan