Instructions that keep it honest

The instructions are the difference between a helpful assistant and a confident liar. They are a short document, and every line does a job.

Who it is: the team it serves and what it covers. What it answers from: only the attached documents, never general knowledge, never guesses. How it answers: short, answer first, cite the document and date, give the next step. What it does when it does not know: say so plainly, and name the person or team to ask. What it never does: give advice on the judgement questions from Module 01, discuss anything outside its scope, or invent a policy. And a line about tone, in the voice of the team.

Then paste in your three model answers from Module 01 as examples. Models follow examples better than rules, and those three set the standard.

In practice. A team's first assistant was built without a "say when you do not know" line. In testing, asked about a policy that did not exist, it wrote one. The line was added. Asked again, it said "I do not have that; ask HR." The lead says that single behaviour is what made the team trust the rest.

Write instructions for a shared team assistant. The team: [ who ]. It covers: [ the version-one question list ]. It answers only from the attached documents. Rules: answer first, in under 120 words; cite the document title and date; give the next step if there is one; if the answer is not in the documents, say "I do not have that" and direct to [ person or team ]; never answer these judgement questions, redirect them instead: [ list ]; never invent or generalise. Tone: [ describe ]. Then include these three example answers as the standard: [ paste ]. Keep the whole thing under 400 words.

At your desk. Write the instructions with the prompt. Paste in your three model answers. Load the instructions and the documents into the assistant.

Write it down.

• My "does not know" line:
• The judgement questions it must redirect:
• Words in my instructions: