Practice — Claude Code: Models, Cost, and Context (17 questions)

For each of the following four tasks, recommend a specific model (Haiku, Sonnet, Opus, or Fable 5) and an effort level (low, medium, high, xhigh, or max), and justify each choice using the capability–cost–latency tradeoffs and effort-level guidance from this subject:

1. A subagent whose only job is to grep a 40-file directory for every call site of a deprecated function and return a plain list of file paths and line numbers — nothing else.
2. An interactive pairing session where a developer is live-debugging a flaky integration test with you, watching the terminal the whole time, and wants to iterate quickly.
3. A root-cause investigation into an intermittent production outage that has resisted three previous debugging attempts by other engineers, where the ambiguity is in what's actually happening, not in how to fix a known cause.
4. A CI job that runs a nightly, fully unattended lint-and-format pass across the repository with no human watching it run.

A developer is deep into a long Claude Code session (several hours, dozens of turns) and realizes their project's CLAUDE.md is missing an important rule: "always run npm run typecheck before considering a task done." They open CLAUDE.md in another editor and add the rule while their Claude Code session keeps running in the terminal.

1. Trace precisely what happens to the prompt cache as a direct result of saving that edit. Does the next turn in the session become more expensive because of it? Explain mechanically why or why not, referencing what's actually part of the cache key.
2. Separately from the cache question, does the running session's behavior change as a result of the edit — will Claude actually run npm run typecheck on its next action because the rule is now in the file? Explain what determines this.
3. Given your answers to (1) and (2), what should the developer actually do if they want the new rule to take effect in the current session, and what's the cheapest way to do it?

A developer sees Claude Code's /fast command and enables it, expecting it to behave like choosing a smaller, cheaper model tier — "faster because it's dumber." Based strictly on how fast mode actually works in Claude Code, which statement correctly describes what turning it on changes?

You've been running Claude Code all morning across three unrelated tasks and want a clean start before your next one.

Which of the following best describes what running /clear actually does?

You're 90 minutes into a session that's about to auto-compact, and you want the resulting summary to keep the list of files you've touched and the API contract you agreed on — not the exploratory dead ends along the way.

What should you do?

Your team is averaging close to $13 per developer per active day on Claude Code, and monthly per-developer spend lands in the $150-250 range.

Based on the benchmarks in this subject, is that alarming?

You leave a Claude Code session open over a long lunch, then send one short follow-up question.

Why can that single short message end up costing far more than a similar message sent earlier in the same session?