*Este artículo fue creado utilizando entrada de voz con IA (Aqua Voice). Tenga en cuenta que puede haber algunas inconsistencias.
Are you still stuck at “being good at prompts”? — AI Evolution and Three Types of Engineering
Hola. ¿Cómo están todos?
Soy Hiro del Laboratorio de Full Body Tracking.
Today, instead of the usual full-body tracking talk, I want to talk about AI.
I use an AI tool called Claude Code for all sorts of things — writing articles, writing novels, generating Blender 3D scripts, and more.
But just using it normally isn’t enough.
Today, I’ll talk about how the way we use AI has evolved, what has changed alongside AI’s evolution, and why I arrived at the way I use it now.
There Are Three Generations of AI Usage
Based on my experience using AI, I feel that the evolution of usage can be broadly divided into three stages.
The first is “Prompt Engineering.”
The second is “Context Engineering.”
The third is “Harness Engineering.”
Let me explain each one.
Generation 1: Prompt Engineering
If you use AI at all, most people probably started here.
There’s a chat screen like LINE, you say “Please do such-and-such,” and it replies “Sure, here’s such-and-such.” That’s the gist of it.
Prompt engineering is the skill of writing those instructions well.
For example, something like this:
“You are a professional editor. Please proofread the following text.”
You give the other party context and prerequisites to draw out more specialized answers. The key is writing good instructions.
This remains important even in later generations. However, it had a fatal weakness.
You Have to Say It Every Single Time
“You are a full-body tracking researcher, your first-person pronoun is ‘I,’ your writing style is casual-polite…”
You have to write this every time, within a limited character count.
AI can’t remember everything, so you need to be as concise as possible, but also accurate.
Honestly, it had its limits.
It Forgets When the Conversation Gets Long
Humans are the same way, right? When you keep talking for a while, you end up thinking “Wait, what were we originally talking about?” AI is the same.
Quality Becomes Inconsistent
As conversations get longer, quality starts to waver. You set it up as “a full-body tracking researcher,” but partway through, it turns into a completely different character.
You Can’t Leverage Past Knowledge
You want to say “Please absorb all the information from articles I’ve written before and use it,” but that’s impossible. There’s a limit to how much AI can hold in its memory at once.
In the end, the approach of “trying to handle everything with prompts alone” started showing its limits.
So the question became: is there another way? And AI usage evolved in a new direction.
Generation 2: Context Engineering
What is context engineering? It’s the skill of designing “what information to show AI, at what timing, and in what order.”
If prompts are about “what to ask,” then context is about “what to show AI.”
Three Things to Do
First, feed it the necessary information.
When having AI write an article, you pre-load relevant past articles, the latest device information, reader feedback, and so on. Not everything — just what’s relevant to the current task.
Next, manage information freshness.
You judge “this information is current” and “that information is outdated,” and prioritize showing the newer stuff. If you have it write “latest recommendations” based on old information, that’s going to be misleading.
Finally, design the order of information.
The order in which you show things to AI changes the output. “First, help it understand the VRChat worldview. Then, help it understand full-body tracking technology. Finally, present the current topic.” Reversing this order produces completely different results.
Four Techniques
Context engineering is said to have four main techniques.
Selection: Choosing the necessary information.
Compression: Condensing information and passing only the key points.
Segmentation: Breaking information into blocks and presenting it in stages.
Writing Out: Removing unnecessary information to reduce noise.
Feeding everything at once actually backfires. The key to context engineering is presenting only what’s needed, in the form it’s needed.
Generation 3: Harness Engineering
The final one to emerge is harness engineering.
When you hear “harness,” some people might think of automotive wiring. But the harness I’m talking about here originally refers to the full set of equipment for controlling a horse. Tools that guide a horse in the right direction.
Harness engineering is the skill of designing the environment itself in which AI operates, so that it always runs in the right direction.
Similarities to Test Harnesses
By the way, since Anthropic — the company that makes Claude Code — is a software company, I think this “harness” shares the same etymology as “test harness.”
A test harness is a software development term for a system that structurally controls the execution environment of a test target and makes results verifiable.
Equipment to make a horse run correctly. A system to verify test targets correctly. Environmental design to make AI work correctly. They’re all connected.
What Am I Actually Doing?
Let me introduce the harness I’ve actually built.
First: Skills (Personas).
Rules like “Hiro writes like this,” “the first-person pronoun is I,” “this is prohibited” are defined in files. AI reads these files and operates from the start with the premise “I am Hiro.”
No need to write “You are Hiro” in the prompt every time. It’s already on that track.
Second: Knowledge Base.
Articles I’ve written in the past, VRChat knowledge, information about full-body tracking devices. All of this is organized in folders, and AI can reference it when needed.
No need to write “Full-body tracking is a technology that follows the movement of the entire body…” in the prompt every time. It knows how to search and how to find the knowledge.
Third: Workflow.
There’s a set procedure for writing articles: confirm the theme, design the hidden theme, outline, write, revise. This procedure is written in a file, so AI can work through “what to think about first” step by step.
Fourth: Memory.
Records like “this person made this decision before” and “this policy was set for this project” are saved in files, and AI can read them in new sessions.
This might be the biggest one. Memory carries over beyond the limits of the context window.
Fifth: Quality Checks (Checkers).
When AI produces output, it doesn’t go straight to publication. A checker runs automatically. “Does it match the original definition?” “Is the first-person pronoun correct?” If something fails, the process goes back to the beginning of the flowchart and redoes it.
Instead of saying “Be careful” in a prompt, checks are built into the system.
The Difference Between “Requests” and “Systems”
The important point here is: “But weren’t there condition definitions like ‘do it this way’ and ‘write it this way’ at the context engineering stage too?”
Yes, there were. But those condition definitions were actually just requests.
Even if you said “Don’t do this” or “Don’t write like that,” there was no system in place to ensure the correct path. Since they were requests, AI could forget or ignore them.
Harness engineering isn’t a request — it’s structure. It lays down rails so AI can only run in the right direction. That’s the fundamental difference.
Comparing All Three Side by Side
Let me organize everything so far all at once.
Prompt Engineering:
The skill of making good “orders” to AI.
In restaurant terms: Placing your order.
Context Engineering:
The skill of giving AI the “right ingredients.”
In restaurant terms: Sourcing the right ingredients.
Harness Engineering:
The skill of designing the “kitchen” where AI works.
In restaurant terms: Designing the kitchen.
They’re completely different, right? But you need all three.
The best order (prompt) x the best ingredients (context) x the best kitchen (harness).
When all three come together, AI can actually cook. If any one is missing, it doesn’t work.
A House Analogy Makes It Even Clearer
For those who didn’t quite connect with the restaurant analogy, here’s a house version.
Harness = Foundation. Once built, it lasts forever.
Context = Walls and pillars (structure). Can be reconfigured per task.
Prompt = Interior design (furniture arrangement). Can be changed every time.
A solid foundation, proper walls, and well-arranged interiors. Only when everything comes together does it become a “house.”
Conversely, no matter how elaborate the interior (prompt), without the foundation (harness), you’re rebuilding from scratch every time. Without walls (context), you don’t even have a room.
Real Example — How I Wrote “The Metaverse Lover”
Where I felt the difference between these three most strongly was with the novel “The Metaverse Lover,” all 12 chapters, which I wrote recently.
What Would Happen with Just Prompts?
“The setting is VRChat, the theme is romance in a virtual world, please write it in 12 chapters.”
Write that, fix what comes out, give more instructions, fix again. Repeat.
Something does come out of this. But what you get is an “AI-ish novel.” The VRChat descriptions are shallow, and the emotion of full-body tracking doesn’t come through.
What I Did
I used all three layers.
At the prompt level:
“This chapter goes something like this.” “This scene should have this kind of atmosphere.” I kept individual instructions simple.
At the context level:
I loaded in the content of previous chapters, VRChat’s newcomer-guide culture, the emotional experience of full-body tracking, and the atmosphere of dance halls. The information this particular chapter needed right now.
At the harness level:
The “Hiro” persona ensured VRChat authenticity. The knowledge base ensured accurate full-body tracking descriptions. Quality checks confirmed “Is there anything unrealistic for VRChat?” and “Is the story coherent?” Memory maintained consistency across all 12 chapters.
The hardest part, honestly, was the context. If you can’t put something into words yourself, you can’t convey it to AI either. Articulating “this particular feeling of VRChat” is genuinely difficult. There were many failures.
But when all three clicked together, I was able to write a 12-chapter long-form work with consistent quality. If any single element had been missing, I don’t think it would have worked.
The Three Aren’t in Opposition — They’re Layers
“So is prompt engineering outdated now?”
Not at all.
I explained that the three belong to different generations, but that doesn’t mean the older ones become unnecessary. You use all of them. They’re just different layers.
Even after building a harness, you still give instructions at the prompt level: “Write this chapter like this.”
Even if the context is well-organized, if the prompt is sloppy, the output will be sloppy.
Only when all three come together does AI perform at its full potential.
Resumen
With prompts alone, you’re a customer placing orders from scratch every time.
When you organize the context, AI gets good ingredients, and the quality of output skyrockets.
Once you build a harness, AI becomes a teammate. It carries your knowledge, works in your style, and runs on your behalf.
I’m a full-body tracking researcher, but it’s the same with full-body tracking.
With just a headset and controllers (= prompts only), you can enter the virtual world.
Choosing the right world (= organizing the context) changes the quality of the experience.
Putting on full-body tracking sensors, calibrating, and setting up the environment (= building the harness) — that’s when you first get the feeling of “being there yourself.”
It’s the same with AI.
Only when you align all three layers does it become “your AI.”
I hope all of you will try the stages that lie beyond prompts.
Hiro
¡Gracias por leer! ¡Síguenos en X y suscríbete a nuestro canal de YouTube!