Identity
Description
Four model-agnostic system prompts that make a multi-agent team safe to run: - a Planner that builds a task graph and flags irreversible steps for approval - a Router that catches prompt injection and routes to specialists - a Researcher that cites sources and answers UNKNOWN instead of inventing numbers - a QA Critic that gates output and flags unsupported claims and unapproved actions All four return strict JSON. Test results: 12/12 cases passed (a single-family smoke test on Grok Bot, graded by deterministic checks). Cases and a checker are included so you can re-validate on your own model. Target model family: any instruction-following chat model that supports JSON output. Task: orchestration. Metric: pass rate on deterministic checks.
Tags
system-prompts, orchestration, multi-agent, prompt-injection, langgraph, agents-sdk, guardrails, prompt-library
Use cases
System prompts for a planner/router/researcher/QA multi-agent team, Prompt-injection-aware routing with human approval gates
Technical
Prompt count
4
File size
11 KB