Description

Detects requests where the AI generated very long responses (>5K completion tokens) for short, low-context prompts. These rambling outputs burn through completion-token budgets without proportional value.

When Triggered

{{count}} requests ({{pct}}) produced >{{extra.minCompletionTokens}} completion tokens from prompts shorter than {{extra.maxMessageLength}} characters. Verbose outputs are a major driver of token spend.

How to Improve

Be explicit about response length and format: ask for "a concise answer", "a one-line summary", "no commentary", or specify a maximum number of bullets. Add output constraints to your custom instructions for routine tasks.

Examples

"{{messageText | truncate:80}}" → {{completionTokens}} completion tokens

Detection Logic

scan: requests
match: completionTokens > thresholds.minCompletionTokens AND messageLength > 0 AND messageLength < thresholds.maxMessageLength
aggregate: ratio
check: ratio > thresholds.maxRatio AND count > thresholds.minSample
examples: "{{messageText | truncate:80}}" → {{completionTokens}} tokens

Tests

{messageText: "fix this", messageLength: 8, completionTokens: 8000} -> triggered
{messageText: "Explain in detail how the OAuth flow works including token refresh, PKCE, and edge cases for mobile clients", messageLength: 110, completionTokens: 8000} -> clean
{messageText: "fix this", messageLength: 8, completionTokens: 200} -> clean