What a harness is

Here is the distinction that makes the rest of this course make sense.

The engine is Claude, or GPT, or Gemini. It reads text and works out what comes next. Genuinely remarkable, and swappable.

The harness is everything else: the instructions, the files it can read, the memory of what you decided, the tools it can reach, the checks before anything counts as done, and who has to approve what. Drawn out, the shape makes the point by itself.

None of those pieces is the model. You built the first pieces of it in part one without being told the word, and gave it context in the last lesson. This lesson writes the part that governs behaviour.

Do not think "I am building around Claude." Think "I am building an environment that Claude currently runs in."

Built the second way, changing model is a settings change. Built the first way, changing model is starting over. Given how many times the best available model has changed hands in the last two years, that is not hypothetical.

When the AI built your AIOS, it wrote an instructions file. Depending on the app it may be called AGENTS.md, CLAUDE.md, or it may be the text sitting in your Project's custom instructions box. Same thing in every case: the file the AI reads first, that tells it how to behave.

Open it. If you cannot find it, ask: show me my instructions file, the one you read before answering. If it turns out not to exist, ask for one to be written from everything it already knows about you.

The AI may also have invented its own name. Asked to build an AI OS from scratch, ChatGPT is fond of something like AI_OS.md. The name changes nothing; the job the file does is what matters. One case where the name does earn its keep: AGENTS.md and CLAUDE.md are conventions that some desktop tools, Claude's and ChatGPT's coding and work modes among them, read automatically when they open a folder. If your AIOS lives in a folder, ask for a copy of your instructions saved under one of those names and the file starts loading itself.

It will have arrived with sensible defaults. Your job is to make two sections yours.

Who you are helping, and what to read first

Already half done from the interview and the context lesson. Check the list of files is right, and that it says to read the relevant ones rather than all of them for a quick question.

What you do not want

The section that is worth more than all the rest put together once you have used the setup for a fortnight. Mine says things like:

• Long preambles. Start with the answer

• Do not tell me the task is a great question

• Three options when I asked for a recommendation. Recommend, then note the alternative

• Padding to look thorough

Yours will be different. Write two now, then add to it every time something irritates you. That list is the single highest-return file in the whole system.

Can you say, in your own words, what the difference is between the model and the harness? If you can, you now have the mental model that most of the AI conversation is missing.

There is one more section of the instructions file to write, and it matters enough to get a lesson of its own: the list of things the AI must never do without you.

Stuck anywhere in this lesson? Reply to any Build Notes email and tell me where. I would rather fix the lesson than have you quietly give up.