Instructions

Instructions are standing directions a product sends an AI with each piece of work. The vendor, the firm or the person using the product may write them. They are text for the AI to follow, not a setting that makes a particular answer impossible.

A principal enters a house rule in the firm's drafting product: "Never describe our restricted advice as independent." Months later, compliance finds "As independent advisers, we reviewed your options" in an annual review letter to Janet Fairlie. The adviser signed and sent it. The product had sent the rule with the drafting request, but it had no separate check for that wording. The instruction did not stop the AI from writing it.

The firm also has a procedures manual and letter templates. Like them, instructions state how work should be done. Fixed template wording stays in place unless someone changes it; an AI may miss an instruction in one draft. The vendor may also send directions the firm cannot see. The firm still needs to check the letter before it goes out, rather than treating a written instruction as a lock.

Not to be confused with

• Calling them the prompt suggests something typed for one piece of work. Instructions go with every piece of work, and the person asking may never see them. Vendors often call them the system prompt.
• Writing instructions is not training the AI. They are sent again with every request and change nothing in the AI itself.
• Instructions, in this entry, are directions to the AI. They are not a client's instructions to the firm, or the firm's instructions to a pension provider.

Overheard

Where it sits

Instructions

• is part ofHarness
• works withModelChat app vs AI inside a productPromptContextPermissionNon-determinism

Grey words are still being written.