Teaching Kids Programming vLog: Harness Engineering — Commanding AI to Get Work Done

Teaching Kids Programming: Videos on Data Structures and Algorithms

This post shares a hands-on programming lesson with my son, using the VS Code AI Agent to clean up his GitHub Pages homepage. It reflects on how programming education is changing in the AI era: beyond writing code manually, kids now need to learn how to describe requirements clearly, command AI agents, understand Git/GitHub workflows, and review AI-generated code responsibly. The post also introduces the idea of Harness Engineering, where the LLM is only the “brain”, while the surrounding tools, workflows, and validation mechanisms are what allow AI to truly get work done.

Last year, my younger son learned some GitHub, Markdown, and HTML. He also created his own personal homepage on GitHub Pages. The page worked, but it was still quite messy: the layout was not very consistent, the content structure was unclear in places, and the code looked a bit like it had been written wherever inspiration happened to strike.

So this time, I used it as an opportunity to show him how to use an AI Agent inside VS Code to improve the website.

In the past, when we taught kids programming, the focus was usually on “how to write code by hand”: how HTML tags work, how to tweak CSS, and how to run Git commands. But things are different now. With AI Agents, a lot of code no longer needs to be written line by line manually. The more important skills become: can you describe the goal clearly, can you command AI to get work done, can you understand what it changed, and can you judge whether the generated code is correct?

This is what I understand as Harness Engineering.

In simple terms:

Harness = Agent – Model (LLM)

The Harness Engineer Post by OpenAI:

https://openai.com/index/harness-engineering/

Redefining the role of the engineer
The lack of hands-on human coding introduced a different kind of engineering work, focused on systems, scaffolding, and leverage.

That means everything around the large language model, except the model itself, can be considered part of the harness. This includes tool calling, file reading and writing, code editing, terminal execution, context management, task decomposition, workflow control, permission boundaries, result validation, and so on.

The LLM is the “brain”, but the harness provides the “hands, feet, and working environment”.

Without a harness, a large model can only chat. With a harness, it can actually start doing work.

The demo I gave my son was very simple: we opened his GitHub Pages project, then gave the Agent a prompt inside VS Code, something like:

“Please help me clean up this personal homepage. Make the page structure clearer and the styling more consistent. Keep the existing content, and don’t make it too complicated.”

The AI Agent then started reading the files, understanding the structure, modifying the HTML and CSS, and even explaining what changes it had made. For a child, this process is very intuitive. Programming is no longer just about burying your head in the code and typing everything yourself. You can also act like a small “project manager”: describe the goal clearly, then let AI help with the implementation.

Of course, this does not mean we no longer need to learn programming.

Quite the opposite. The more AI we have, the more important the fundamentals become. Otherwise, if AI writes something wrong, you won’t notice. If it messes up the project structure, you won’t know how to recover. If problematic code gets committed, you won’t know where the issue is.

So along the way, I also taught my son some basic GitHub workflow commands:

git add

Add modified files to the staging area.

git commit

Create a clear record for this set of changes.

git push

Push local changes to GitHub.

We also talked about Pull Requests, or PRs. Even for a small personal project, it is a good habit to use PRs to review changes, describe what has been modified, and check the code before merging.

I think this is the kind of programming education kids really need in the AI era.

It is not about memorising every syntax rule, nor about writing HTML just for the sake of writing HTML. It is about understanding how a software project is modified, tested, submitted, and maintained. AI can help you write code, but you are still responsible for defining the requirements, checking the result, managing versions, and judging quality.

In the future, many jobs may work like this: you may not personally handle every tiny detail, but you need to know how to break down complex tasks, how to direct AI effectively, and how to take responsibility for the final result.

That is also one of the ideas I wanted to share with my son through this vLog:

Knowing how to write code is important, but knowing how to use tools to solve problems is even more important.

AI Agents do not replace learning. Instead, they move learning up one level. In the past, children learned “how to write code”. Now, they also need to learn “how to make AI write reliable code”.

And that is actually quite interesting.

My son originally just wanted to improve his personal homepage. But in the process, he experienced several key parts of real software engineering: describing requirements, collaborating with AI, modifying code, committing with Git, creating Pull Requests, and reviewing code.

That is much closer to real-world engineering practice than simply memorising a few HTML tags.

AI/Prompt Engineering

• Teaching Kids Programming vLog: Harness Engineering — Commanding AI to Get Work Done
• AI Coding Agents are Reshaping the Barrier to Programming
• Prompt Engineering: Sharing a Prompt for Preparing Your Exams
• ChatGPT: New Image Creation (art school)

Prompts

• ChatGPT Image 2.0: Casual Selfie of Me
• A few ghibli images that chatgpt creates recently (2026-03-19) | blurt
• AI Transforms Photos into Ghibli Style: ChatGPT 4.0 Costs Money, but Here’s a Free Method

steemit：Teaching Kids Programming vLog – Harness Engineering – Using Visual Studio AI Agent

–EOF (The Ultimate Computing & Technology Blog) —