Tool instructions

• You expose multiple tools and need the model to pick the right one reliably.
• Tool misuse is causing latency spikes or rate-limit hits.
• You’re adding a new integration and must protect PII or billing data.

Tool instructions control cost and risk. Good instructions reduce wasted tool calls, shrink latency, and avoid policy violations. PMs decide the priority order of tools, what errors to surface, and how to recover when a tool fails—directly shaping user trust and SLOs.

Document each tool with intent, required fields, and when NOT to call it. Provide short, contrasting examples. Add budgets: max calls, max parallelism, and fallback behavior. In 2026, pair tool instructions with runtime scoring (success rate, retries, cost per call) and auto-disable tools that dip below thresholds until fixed.

A sales copilot has CRM search, email send, and calendar tools. By adding a rule to prefer search before composing emails and capping retries at two, tool success rises to 94% and median latency drops from 2.4 s to 1.5 s without losing deal coverage.

• Letting the model guess tool parameters, leading to partial or failed calls.
• Not defining negative cases (when to avoid a tool), causing loops.
• Ignoring rate limits, which triggers throttling and poor UX.