Harness, Part 3: The System Prompt
Your agent gets an identity - your answers become a system prompt that lives in its own file.
What this is about
So far your agent talks the way the model happens to talk: pleasant, generic, interchangeable. What's missing is the first message you never see - the system prompt, stage 3 of the staged model. It stands before everything else in every call and shapes it all: tone, task, stance, limits. Invisible to the user, the constitution for the model.
The decisive move of this part: the identity goes into its own file, not into the code. Whoever wants to change their agent's character should get to edit a text - not have to program. Your coding agent will ask you about tone, task and taboos and draft the file from your answers; you read it back and approve. It is your identity, not its.
At the end of this part, this works: the same agent, but unmistakably yours - and you can change that any time with a text editor.
How the build works
As before: copy the prompt, hand it to your coding agent, answer its questions. The prompt first reads harness-doku.html and checks the state from Part 2.
This section belongs to the Member tier. You get access to full code, deeper explanations, and rabbit-hole material.
See tiers →It works when
The same question sounds different with the identity than without - in tone, in focus, in length. And the identity sits as a text file in your folder: edit, save, restart, done. Your agent can now talk the way you want - but it still can't do anything. Part 4: Tools & Loop changes that.