The Year We Didn’t Write Code

A short memoir from the Age of Prompt Gods

That evening, I was sitting in a rocking chair at a retirement home.

Beside me, an AI caregiver was installing my 42nd personality patch.

My grandson suddenly looked up from his holographic tablet, visibly disturbed.

“Grandpa,” he asked, “were programmers in 2025… superstitious?”

I paused.

“What makes you say that?”

He turned the projection toward me. On the screen was a directory I instantly recognized — one of my old projects.

. 
├── src/ 
├── .claude 
├── CLAUDE.md 
├── .cursorrules 
├── .github/ 
│   └── copilot-instructions.md 
├── AGENTS.md 
├── openspec.yaml 
├── prompts/ 
│   ├── system.txt 
│   └── personality.md 
└── README.md

He frowned.

“The core logic is three lines,” he said.
“But the instructions for AI are longer than Sapiens.”

I sighed.

“Those weren’t instructions,” I told him. “They were prayers.”

The Polytheistic IDE Era

“You see,” I began, “we lived through the Great Melee of Language Model Gods.”

• Cursor was one god.
• Claude was another.
• Copilot belonged to some Nordic pantheon.
• Open-source agents were folk deities — unstable, powerful, and poorly documented.

And all of them claimed to follow standards.

If you wrote an AGENTS.MD, Cursor would say:

“I don’t accept this format.”

You renamed it to .cursorrules, and Claude replied:

“Sorry, I’m blind in IDE mode.”

You unified everything into openspec.yaml, and all AIs agreed:

“This looks great.
We don’t support it.”

How to Talk to AI (Circa 2025)

My grandson interrupted.

“Couldn’t you just write one file?”

I laughed.

“You don’t understand,” I said. “Each AI demanded to be respected individually.”

To Cursor, you had to say:

“You are a world-class software engineer.”

To Claude:

“You are cautious, humble, and principled.”

To Copilot:

“Give me code. No explanations.”

To open-source agents, you first had to explain:

• what year it was
• who you were
• who they were
• why either of you existed

And every one of these messages had to live in a different file, with a different name, in a different directory.

Otherwise, the AI would get… upset.

Not visibly.
Just enough to:

• Confidently write incorrect code
• Answer B when you asked A
• Hallucinate with moral authority

Debugging the Gods

“How did you debug them?” my grandson asked.

I stayed quiet for a long time.

“With rituals,” I finally said.

Delete .claude, and Cursor suddenly became brilliant.
Add “think step by step,” and Copilot froze.
Change nothing. Restart the IDE. Everything worked.

We called it:

Prompt cache misalignment

Which really meant:

We had no idea why.

A Classic Commit from the Era

I pulled up an old repository.

feat: adjust agent prompts

0 lines of business logic changed
+300 lines of prompts
-200 lines of promptsFixed one bug
Introduced two new ones

My grandson stared at it.

“So… what did you do all day?”

I looked toward the ruins of an abandoned data center.

“We weren’t writing code,” I said.

“We were negotiating with different vendors’ AIs
about how they should interpret the same code.”

The Final Standardization

“What happened later?” he asked.

“Eventually,” I said, “the industry unified.”

His eyes lit up.

“Really?”

“Yes,” I nodded.
“Everyone standardized on… their own standard.”

He fell silent.

“It was like the browser wars,” I added,
“except this time we never rendered a page.

We just wrote instructions for the browser.”

Epilogue (from the README)

I read him the last line of the README:

If you open this project on the same day, in the same mood,
using the same IDE, the same AI, and the same version,
it should work.

He nodded slowly.

“Now I understand,” 
he said
“why .claude files were eventually banned by law.”

I smiled.

A genuine smile.
Because in that moment, I knew:
Humanity had finally graduated from prompt engineering.

And moved on to something else. We didn’t understand yet.