ð This time, we are at the third stage: harness engineering. Even if prompts and context are well-prepared, you cannot entrust tasks to AI with peace of mind unless there is a mechanism that allows it to operate safely and with reproducibility. We will explore this by adapting it to English learning.
Introduction
ð Erica: “We have prepared both the prompts and the context. But there is still something missing. That is the harness we are discussing today. Let’s look together at the concept of building ‘safety fences’ around the AI.”
A harness is also a safety device used in mountain climbing. In AI engineering, the term ‘harness’ carries exactly that same image.
ð What is Harness Engineering?
ð¹ Rules on the environment side where the AI operates
Harness engineering is about creating a scaffold where AI can work while being verified safely and with reproducibility. Asking the AI to ‘write tests’ falls within the scope of prompting, but preparing the actual mechanism that says ‘you cannot proceed to the next step unless the tests pass’ is what a harness is.
ð¹ Agreements outside the prompt
In other words, a harness is an agreement on the operating environment side that does not appear in the instructions given to the AI. Rules such as ‘this data cannot be deleted’ or ‘you cannot go outside this range’ are prepared in advance outside of the AI.
ð Adapting the analogy of secretarial work to English learning
ð¹ A scaffold that says ‘no appointments after 5:00 PM’
In a scenario where you entrust schedule management to AI, if there is a mechanism that ‘rejects any attempt to schedule something after 5:00 PM,’ you can entrust the task with peace of mind even if the AI acts somewhat freely. This is the practical image of a harness.
ð¹ In English learning: ‘You cannot proceed until you meet today’s goal’
Adapting this to English learning, a mechanism that says ‘you cannot move on to new vocabulary until you finish today’s quota (e.g., reviewing 10 words)’ acts as a harness. Deciding on a rule that you must read through the AI’s corrections yourself at least once before moving on is based on the same idea.
ð Creating your own simple harness
ð¹ Creating a checklist
You can incorporate the harness concept without building an elaborate system. Simply keeping a checklist for yourself, such as ‘I will not accept any corrections that do not meet these three criteria,’ is a perfectly valid simple harness.
ð¹ Instructions like ‘Ask me if this is not met’
Always including a sentence in your instructions to the AI such as, ‘If there is insufficient information, please ask me first.’ This is also a type of harnessâa mechanism that stops the AI from running wild.
Conclusion
ð Erica: “Next time, we finally move on to loop engineering. It’s about the mechanism for running AI on your behalf. From here on out, there’s a Japanese-English prompt set using today’s content, as well as an exclusive, cute, and sassy bromide image of Erica (vertical) to cheer you on, so be sure to check it out. It’s sized perfectly for a smartphone wallpaper.”
From here on is a set of ‘Japanese Prompt x English Prompt + Exclusive Erica Image’ that you can use exactly as is. For the sake of your English studies, please take a look at both prompts together.
â¼ Paid content starts here (200 yen) â¼
