AI / LLM Integration

Tool calls: free-recall review

Crux Free-recall prompts across the tool-calls unit. Answer each in your own words first, then reveal the model answer and compare.

Your altitude — climbing toward senior

ZeroJuniorMiddleSenior

You are at senior altitude — in orbit

◷ 14 min

Retrieval beats re-reading. For each prompt, say or write a full answer from memory before you open the model answer — the effort of recall is what makes the mechanism stick.

Goal

Reconstruct the unit’s core mechanisms — the round-trip contract, the validation trust boundary, the runaway guard, parallelism, tool_choice, and the schema token cost — without looking back at the lesson.

Recall before you leave

01
Describe the tool-use loop. Who is the caller, who is the callee, and what makes each iteration expensive?
02
Why must you validate tool arguments, and what does 'validate' actually mean for a mutating endpoint like cancel_order?
03
Why is an unguarded tool-use loop both a correctness risk and a cost risk, and what two guards do you add?
04
When can the model parallelize tool calls, and when is a chain inherently serial?
05
Explain the four tool_choice settings and when a senior reaches for each.
06
Where does the token cost of tools actually come from, and what is the standard mitigation?

Recap

If you could reconstruct each answer from memory, you hold the unit’s spine: the model is the caller and your code the callee, so every round trip is a fresh, transcript-re-sending model call; arguments are untrusted input you schema-validate then authorize and existence-check, returning errors as tool_result; an unguarded loop runs away, so cap iterations and detect repeats; parallelism helps only independent calls; tool_choice (auto/any/tool/none) sets whether and which tool fires; and the re-sent tools array is the token cost prompt caching exists to absorb.

Continue the climb ↑Tool calls: code and schema reading