passdrill
AI & LLM Engineering · AI Agents & Tool Use · Card 031/032 hard

A customer-facing chatbot uses two separate, simultaneously-running LLM calls on every incoming user message: one call's only job is to screen the message for inappropriate or out-of-policy content, while a completely separate call independently drafts the actual response to the user's question; the two calls' outputs are then combined programmatically before anything is shown to the user. Why does Anthropic's guidance describe handling guardrails this way, rather than having the single call that drafts the response also perform its own content screening?

  1. Because this is an application of the parallelization pattern's sectioning variation, and separating the screening concern into its own dedicated call tends to perform better than having one call juggle both the core response and the guardrail check at the same time
  2. Because a single LLM call is technically incapable of ever producing more than one type of output, so screening and drafting could never both happen within one call regardless of design
  3. Because running two calls is always cheaper than running one call, so splitting any task into two simultaneous calls reduces cost regardless of what each call does
  4. Because Anthropic's documentation states that content screening must legally be performed by a human reviewer, and no LLM call of any kind is permitted to perform it
Next card → Shuffle