Almost nobody writes prompts that other people will want to use.
Most prompts in 2026 still get treated the way we used to treat search queries — dashed off, semi-grammatical, full of personal shorthand. "ok but make it more like before but for the case where" — that kind of thing. The model figures it out. You get an answer. You close the tab.
Fine. For one-off prompts, that's correct. The problem is that prompts increasingly aren't one-off.
They get checked into repos. They get shared in Slack and reused. They get tuned. Versioned. Forked. They become the load-bearing instructions for agentic systems where the model is calling tools, making decisions, and writing code that ends up in production.
A prompt that ships to production is code. You do not write production code the way you write a Slack message to your past self.
The habit
When I'm writing a prompt that's going anywhere except into the next thirty seconds of my own life, I write it as if I'm about to hand it to a smart colleague who just walked into the room. They don't know what I'm trying to do, why I'm trying to do it, or what's already been tried.
That single move — smart colleague, just walked in, no context — does three things almost for free:
- It forces me to state the goal, not just the request.
- It forces me to name the things I'd otherwise let live in my head.
- It forces me to specify what done looks like.
The output improves dramatically. Almost incidentally.
The test
Take a prompt you wrote in the last week. Now imagine handing it to the most senior engineer you know — somebody who's good but tired, who's about to give you fifteen minutes — and ask: would they know what I want?
If the answer is "they could probably figure it out," you wrote a Slack message. Rewrite it.
If the answer is "yes, instantly," that's a prompt. Save it.