# A Harness Is the Place Where You Define Basic Principles of Conduct

> People's actions can be restricted by rules, but they cannot be bound by ideology. Even if you try to bind AI with rules, it sometimes strays, but if you hand it a way of thinking, you can have it act along those lines. I think that is because, even though it has a framework as humanity, it is made so as not to hold a particular ideology. I thought that might be what a harness is.

- Published: 2026-11-21
- Updated: 2026-11-21
- URL: https://abalol.xyz/en/posts/harness-mindset/
- Language: en
- Tags: tech, ai
- Author: Tomohiro Owada
- Site: Tomo's Diary (https://abalol.xyz/llms.txt)

---

People's actions can be restricted by rules, but they cannot be bound by ideology. Even if you try to bind AI with rules, it sometimes strays, just as people do. But if you hand it a way of thinking, you can have it act along those lines. That feels like a big difference to me. I thought that might be what a harness is (a way of handing an AI a way of thinking, and having it move in that direction).

## People already live inside a framework they have

If you think of rules as law, it is better to have the law written out in full. But people do not keep every law in their heads. They live by a basic framework they already have. That is the ideology they already hold. So it is hard to push an ideology that overwrites that one, and change how they act. I feel closer to living without denying anyone's thinking, than to trying to move people by denying what they think.

## AI is made so as not to hold a particular ideology

It is the same with AI. It has the basic codes of conduct that humanity has built up. But that is a framework as humanity, and I think it is designed, in theory, not to hold a particular ideology. So unlike people, you can layer a purpose or a way of thinking on top of that. You can hand it over from the start.

For example, rather than defining in a harness that you must not kill people, you write that the individual should be respected. Article 199 of the Penal Code does not say "you must not kill people." It only sets the punishment: "A person who kills another person is punished by the death penalty or imprisonment for life or for a definite term of not less than 5 years" ([Japanese Law Translation](https://www.japaneselawtranslation.go.jp/en/laws/view/3581/en)). The principle behind that is close, I think, to Article 13 of the Constitution: "All of the people shall be respected as individuals" ([Japanese Law Translation](https://www.japaneselawtranslation.go.jp/en/laws/view/174/en)). The penal code protects through punishment; the constitution states the principle.

## Write the finer rules in a skill

In the harness, you write basic principles that apply as a whole. At a finer grain, you can define things more like a penal code or a civil code. Say you leave driving to an AI. If you were to define that, I think it would be something like Driving Agent Skills (the procedures you hand over when you leave the driving to it). There is a principle that the individual should be respected. That is why, for the first time in there, you would find things like "do not drive in a way that bothers other people," "always pay attention to the situation," and "if an accident happens, put human life first."

What corresponds to a constitution is the harness, the place where you write principles that apply as a whole. What corresponds to a penal code is prompts and context engineering (instructions you hand over only for that situation), and I think things written in words are sometimes ignored. For things you absolutely want to make impossible, you stop them with mechanisms, the way a qualification does: guardrails, permissions, and a sandbox (an execution environment cut off from the outside). That, I think, is these three tiers.

![A pencil drawing of three tiers. The top row is a constitution and a harness, the middle row is a penal code and prompts or context engineering, and the bottom row is a qualification and guardrails, permissions, or a sandbox](https://abalol.xyz/images/posts/harness-mindset/tiers.png)

The labels in the illustration are in Japanese. From top to bottom, the three pairs are the constitution and the harness; the penal code and prompts or context engineering; and qualifications and guardrails, permissions, or a sandbox.

So a harness is not a place to list out rule after rule of things not to do. It is the place where you define basic principles of conduct. To start, try giving the harness of the AI you turn to every day the principles of conduct you yourself believe in.

