Lesson 2: Steering the agent with a written prompt
21 min read·13 Oct 2026
Module 2 · Lesson 2 · About 45 minutes · Needs: Lesson 1 (rule checks in code, the OUTCOME: line, R-1002)The situation
It is Tuesday 14 April 2026, week 1. Lesson 1 showed what the model can be trusted with on one reply. Now Priya wants the agent's instructions written down: what it is for, what it must leave alone, and when it must stop. Dev asks the question that follows: where does that prompt live, and who can change it?
You answer both with one file, prompts/system_v1.md, and you test it on Tom Becker's R-1002 again: order 7120, the GlowLine smart bulbs, 69 days after delivery. The right outcome is still declined. This time the model works under an agent's prompt, with eight tools on offer.