What makes tool descriptions misleading to a model even when they read well to humans?

Instruction: Explain why human-readable tool descriptions can still mislead an AI system.

Context: Checks whether the candidate can explain the core concept clearly and connect it to real production decisions. Explain why human-readable tool descriptions can still mislead an AI system.

Updated

Example Answer

A description misleads when it promises more than the tool does or hides a decision the caller must make. “Fix customer accounts” might actually mean replacing one billing-address field; the model could infer broader repair authority from the wording.

I'd describe the exact action, prerequisites, exclusions, and returned outcome. I'd also make overlapping descriptions explain a real distinction, such as search by name versus fetch by known ID. Examples should include a request the tool cannot fulfill. Tool descriptions are guidance, not permission grants; the backend enforces the real boundary. I'd replay confusing user intents and inspect selection, because polished prose can still leave the model with two equally plausible choices.

Why this works

Uses overpromising and overlap examples, then tests their effect on selection.

Practice follow-up

How would you rewrite a tool described only as “manage documents”?

Your preparation path

Work through these questions in order. Read the answer aloud, then explain it in your own words.

1. Start with the foundations

Build the vocabulary and explain the core decisions.

  1. When should a model call a tool instead of answering from context? Free sample
  2. How would you explain MCP to a product engineer in one minute? Free sample
  3. What information should the orchestrator validate before executing a tool call? Free sample

2. Apply it to a real workflow

Practice diagnosis, validation, and everyday tradeoffs.

  1. Why do idempotency and retries matter more with AI-driven tool use? Free sample
  2. How do you keep permission boundaries clear when many tools share overlapping capabilities? Member answer
  3. A user revokes access to a connected app mid-session. How should the AI workflow behave? Member answer

3. Prepare for senior discussions

Explain failure boundaries, recovery, and production choices.

  1. Design a fallback strategy when tools are unavailable, degraded, or unauthorized. Member answer
  2. How do you implement MCP step-up authorization without losing scopes or retrying forever? Member answer
  3. An MCP request resumes after user confirmation. How would you validate its requestState? Member answer

Related Questions