"Write for machines," they said. So I did. Turns out I was trying too hard.

I've always argued we need to write for machines in the age of AI. Structure matters. Organization matters. Clarity matters.

But here's where I might be wrong: I confused "writing for machines" with "writing like a machine."

When someone asked if my structured prompts used JSON (programmer-speak for rigid data formatting), I realized I needed to test my own assumptions.

So I took the same prompt and tested four versions:

1️⃣ Natural structure (like clean meeting notes): 64 seconds

2️⃣ Tagged structure (with [SECTION] markers): 83 seconds

3️⃣ JSON (maximum machine format): 97 seconds

4️⃣ Conversational rambling: 120 seconds

The "most machine-readable" format performed worst. Not just slower—the output was mechanical, lifeless, bureaucratic.

🔑 Writing for machines doesn't mean abandoning human organization. It means embracing it.

Why? Because AI models learned from millions of human documents. They're fluent in human organization patterns—clear headings, logical flow, structured thoughts.

They struggle with both chaos AND rigid syntax.

Think about it: What works for a machine that learned from humans?

▶︎ Clear structure ✓ (like good documentation)

▶︎ Logical organization ✓ (like effective emails)

▶︎ Natural hierarchy ✓ (like organized notes)

▶︎ JSON formatting ✗ (like... nothing humans actually write)

The sweet spot is writing that would make both your colleague AND an AI understand immediately. Organized for clarity, not complicated for complexity's sake.

Microsoft researchers confirm: format alone can swing AI performance by 40%. But the pattern holds—natural human structure consistently wins.

Yes, we need to write for machines, but we need to be careful about what that means.

Good structure serves everyone. Over-engineering serves no one.

What "optimization" have you added that actually made things worse?

🔗 Link to research in the comments.