From Prompt to Harness — When AI Giants Start "Harvesting" Startups, Where's Engineers' Moat?

"Harness Engineering" exploded in tech circles recently; many ask what new buzzword this is.

From my software engineer experience, this isn't old wine in new bottles but inevitable evolution. Past two years, we moved from Prompt Engineering (how to communicate clearly) to Context Engineering (providing right data), but now we discover: models are smart enough; real bottleneck returned to architecture itself.

I. Awakening Expression vs. Building Operating Mechanisms¶

I saw Li Hong-yi's research mentioning: Fine-tuning essentially doesn't teach models new knowledge but how to "express." I think Prompt is similar—awakening model's expressiveness.

But diminishing returns are obvious.

You've experienced: no matter how you adjust prompts, models still miss details or "forget" in long conversations. Like human brains handed 100 simultaneous tasks—guaranteed crash. We need Harness—like actual harnesses/constraint devices: breaking big tasks into model-focusable subtasks, letting it interact with real world.

My practice: instead of ultra-long prompts, first "align outlines" with models until fine enough for fresh engineers understanding, then hand to tools like Vibe Code for implementation. This "decomposition ability" is the 10x efficiency multiplier.

II. Human Core Value Getting Pulled Higher-Dimensional¶

With AI intervening more, after extracting ourselves from "coding, reading docs" tedium—will we lose jobs?

Opposite. I believe human core value shifts toward:

• Integration Ability (PM): Define "success" standards.
• Structured Decomposition: Convert fuzzy business needs into AI-executable SOPs.
• Domain Knowledge: Become translators between specialist fields and AI.

That's why Harness progresses fastest in software—these professionals excel "breaking big problems into small ones."

III. Technology Paradigm Replay: Startups' Moat Where?¶

Watching AI giants (OpenAI, Claude) moves, it's like watching old Intel vertical integration vs. TSMC foundry battles replayed.

When Claude's "advisor mode" uses smart models guiding cost-effective models completing tasks, when giants constantly integrate "plugin features" startups built into native products—pure software startups' moats are razor-thin.

Future has two paths:

1. Deep Waters: Entering law, finance, semiconductors—hyper-secretive, data-walled industries. Vertical applications giants can't access yet are real bones.

2. Offline and On-Premise: While everyone fights API ranking, AI running stable disconnected or enterprise private clouds represents true safety.

Conclusion¶

AI competition's core challenge shifts from "making models smarter" to "making models work stably in real world."

If still researching "magic prompts," my advice: invest in Harness architecture research, thinking how building systems auto-verifying, correcting with state memory.

Because determining whether AI lands depends never on models but on constraint-systems running them.

What do you think? In your field, what "deep waters" can't AI giants cross soon?