From Prompt to Harness: As AI Giants Begin 'Harvesting' Startups, Where Is the Engineer's Moat?

The term "Harness Engineering" has been exploding in the community lately, and a lot of people have been asking me: what is this new buzzword, exactly?

From where I stand as a software engineer, this isn't old wine in a new bottle — it's an inevitable evolution. Over the past two years, we moved from Prompt Engineering (how to say things clearly) to Context Engineering (how to feed the right data). But now we've discovered that the models are smart enough, and the real bottleneck has swung back to architecture itself.

1. Awakening Expressiveness vs. Building an Operating System¶

I once came across research by Professor Hung-yi Lee that noted: fine-tuning doesn't fundamentally teach a model new knowledge — it teaches the model how to express what it already knows. I think prompting works the same way: it awakens the model's expressive capacity.

But that has obvious diminishing returns.

You've surely experienced this yourself: no matter how carefully you craft the prompt, the model still drops details, or starts "forgetting" midway through a long conversation. It's like the human brain — dump a hundred tasks on it at once and it will crash. What we need at this point is a system that works like a harness — something that breaks a large task into smaller ones the model can actually focus on, and lets it interact meaningfully with the real world.

My own practice has become: instead of writing one enormous prompt, I first align on an outline with the model — until that outline is detailed enough that a junior engineer fresh out of school could follow it — and then hand it off to a tool like Vibe Code for implementation. That capacity to decompose, not the prompt itself, is what drives a 10× efficiency gain.

2. Human Core Value Is Being Pulled to a Higher Dimension¶

As AI intervenes more frequently and we're freed from the grind of "writing code and reading docs," will we lose our jobs?

Quite the opposite. I believe human core value will shift toward:

• Synthesis (PM-level thinking): Defining what "success" actually means.
• Structured decomposition: Translating vague business requirements into SOPs that AI can execute.
• Domain expertise: Becoming the interpreter between a specialized field and AI.

This is exactly why Harness Engineering is advancing fastest in software development — because people here are already best at turning big problems into small ones.

3. A Replay of Tech Paradigms: Where Is the Startup's Moat?¶

Watching how the AI giants (OpenAI, Claude) are moving right now feels exactly like watching the old battle between Intel's vertical integration and TSMC's foundry model.

When Claude rolls out an "advisor mode" where a smarter model steers a more cost-efficient model to complete tasks, when these giants keep absorbing add-on features that startups used to own and baking them into native products — the moat for pure-software startups becomes razor-thin.

Only two paths remain viable:

1. Go deep into protected waters: Enter industries like law, finance, and semiconductors — domains with extreme confidentiality requirements and real data walls. These are the "vertical applications" the giants can't easily swallow.

2. Go offline and on-premise: While everyone races for the top spot in the API-connected world, AI that can run stably in a corporate private cloud or on edge devices — disconnected from the internet — is where the real staying power lies.

Closing Thought¶

The core challenge in the AI race is shifting from "making models smarter" to "making models work reliably in the real world."

If you're still hunting for that magic prompt, take this advice: invest your energy in studying Harness architecture — think about how to build a system that can automatically verify, self-correct, and maintain state memory.

Because what determines whether AI can actually land in production has never been the model itself. It has always been the operating system that constrains it.

What do you think? In your professional domain, which "deep waters" are even the AI giants unable to cross for now?