Foundations domain 5 · Context management and reliability
- Keep exact facts out of summaries. Amounts, dates and commitments go in a structured case-facts block that's carried into every turn. Summaries blur them.
- Trim tool output to the fields the task needs. Put key findings first, with clear headings; material buried in the middle of long input gets missed.
- Long investigations: use subagents for verbose exploration, keep findings in a notes file,
/compactwith instructions on what to keep, and/clearbetween unrelated tasks. - Escalation: honour an explicit request for a human; escalate policy gaps rather than improvising; ask a clarifying question when a lookup matches several customers. Sentiment alone isn't a reliable trigger. A handoff includes a structured summary, because the human doesn't have the transcript.
- Human review: route by risk and confidence, keep a random sample of the confident cases, measure accuracy per document type and per field, and calibrate confidence against labelled data.
- Provenance and uncertainty: keep claims attached to their sources and dates; report conflicting figures with attribution instead of picking one; say where coverage is thin.
- Crash recovery: save progress as structured state; make side effects safe to repeat.
See also: Professional domain 2 · Context windows and token usage, Foundations domain 1 · Subagents and coordination, Foundations domain 3 · Commands, by symptom, Professional domain 5 · Human in the loop and Professional domain 4 · Diagnosing problems.
Worth knowing (lightly examined)
- Prompt caching matches a prefix: stable content first (tools, then system, then messages), variable content last. Up to 4 breakpoints; the default lifetime is 5 minutes, with a 1-hour option. The exam guide lists caching details beyond knowing it exists as out of scope.
- Errors: 429 (rate limited), 529 (overloaded) and 5xx server errors are worth retrying with backoff, honouring
retry-after. 400, 401, 403, 404 and 413 aren't. The SDKs retry twice by default.
See also: Professional domain 2 · Prompt reuse: caching, modular prompts and Skills.
Exam guide vs todayThinking: manual extended thinking with
budget_tokens (minimum 1,024, below max_tokens) is rejected on the newest models. They use adaptive thinking with an effort setting. With tools, thinking blocks must be passed back unmodified. Caching minimums vary by model (512 tokens on the newest, up to 4,096 on some older ones). Treat any single "1,024 tokens" figure as out of date.See also: Professional domain 2 · Choosing a model.
See also: Exam guide vs today · Foundations (CCAR-F).
Docs: Prompt cachingThinkingExtended thinkingErrors
Practise Domain 5