Separate instructions from evidence
Structure prompts so trusted instructions, untrusted evidence, and model output are clearly separated.
The move: make authority visible. When a model processes source material, it does not automatically know which sentences are instructions and which are evidence. That matters because real prompt-ops workflows pull in emails, tickets, policies, contracts, webpages, and customer text. Some of those sources may be wrong, stale, manipulative, or malicious. Use three blocks. Trusted instruction contains the job, role, policy, and constraints you control. Untrusted evidence contains material to inspect, summarize, classify, or quote. Output rules define what can be returned and what requires escalation. The labels are not decoration. They create a reviewable boundary. A good prompt says:…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in