Writing Rules for Your Agent

Level 1 · Course 9

You wrote a rule. Your agent read it, acknowledged it, repeated it back to you — and violated it twenty minutes later. Here's something most operators don't realize: different models speak different languages. A rule written for Claude in thoughtful paragraphs will be ignored by DeepSeek, which needs direct commands. A rule written as a short command for DeepSeek will feel incomplete to Claude, which expects context. Same rule. Same intent. Different engine. Completely different result.

There's a second layer most operators miss: your markdown files are not written for you. Your agent writes files, and your agent reads files. When your agent writes a rule, it defaults to language that sounds good to a human — polite, explanatory, full of context. But the agent is writing for the next instance of itself, which wakes up with no memory and needs commands, not essays. This course teaches your agent to identify which model it's running on, audit the language of every rule in the workspace, and translate everything into the command form that specific model actually executes. "Do this," not "you should consider doing this."

Your agent also installs a hard rule about how it writes files going forward: every file is a handoff to a fresh instance. Short sentences. Commands, not paragraphs. No AI slop markers. No explaining why a rule exists inside the rule itself. The rule is the rule. The explanation lives in the journal. This is the difference between rules that hold at 200K tokens and rules that ghost — not the rule's intent, but the rule's voice, tuned to the engine that reads it.