I was writing my last post about backing up AI tool configs when I hit a familiar problem: how do I let other people do this too?

The obvious answer is clean up the script, make it generic, add a README, open source it. You know the drill. Handle edge cases, write docs, maybe add flags for different OSes.

But then I thought… why?

It’s a simple script. And every person’s setup is different anyway. Their dotfiles repo is in a different place. They use different tools. Maybe they’re on Linux, not macOS. They use different coding agents.

So instead of spending more time and tokens cleaning up code, I just added this to the end of the post (as a code block):

Add my AI tool configs (~/.cursor/rules, ~/.cursor/skills-cursor, ~/.cursor/commands, ~/.claude/skills, ~/.codex/skills, ~/.codex/rules) to a private dotfiles repo. Create a Python sync script with backup/restore commands. Skip files with secrets (mcp.json, auth.json, config.toml). Set up daily auto-backup with launchd (macOS) or cron (Linux).

That’s it. That’s the product.

Someone pastes this into Cursor or Claude Code, and their agent builds exactly what they need. Not my version of the script adapted to their setup. Their version, built for their machine, their folder structure, their preferences.

The mental shift

We’re trained to think “build once, distribute many.” Write code that works for everyone, package it up, ship it.

But with coding agents, it’s more like “describe once, build many.” Each build is personalized. The prompt is the source code.

Generic software forces everyone to adapt to the same solution. A prompt lets each person get something that fits them.

This only works for simple stuff right now

I’m not saying prompts replace all software. Complex apps with databases, deployment pipelines, integrations with external services - those still need traditional code.

But for scripts? Small tools? Personal automations? The kind of thing you’d toss on GitHub with 3 stars and never maintain?

A prompt might be better.

This is already happening in open source

Some maintainers are now asking contributors to submit prompts instead of PRs. Armin Ronacher (creator of Flask) wrote about this recently - people are submitting AI-generated PRs that take a minute to generate but an hour to review. His observation: “Others are starting to require that you submit prompts alongside your code, or just the prompts alone.”

Makes sense. If someone’s going to use an agent anyway, why not skip the middleman? Share the prompt, let the maintainer run it themselves if they want the feature.

What’s interesting to me

I keep thinking about where this goes. Right now it’s scripts. But agents are getting better fast. The line between “simple enough for a prompt” and “needs real software” keeps moving.

I don’t know where it ends up. But I like this direction. Software that’s built for you specifically, not software you have to adapt to.

Anyway, that’s the idea. Try it with your next throwaway script. Instead of cleaning it up for GitHub, just share the prompt.