Prompt Teardown is a newsletter about prompts. Every issue gives you a few prompts that work, shows you why each line is in there, and tells you the one thing the prompt still gets wrong.
Most weeks that’s one issue. Some weeks two. Some issues carry 3 prompts, some just 1. It depends on what’s worth sending.
It’s model-agnostic. We cover Claude, GPT, Gemini, and others. The patterns apply everywhere.
Each issue is a 4-minute read. Here’s what’s in it and how to get the most out of it.
What you’ll find in each issue
Every issue is built around one kind of problem. A doctor visit you’re walking into cold. A move you haven’t priced. A card offer you can’t read.
Each prompt comes with the same 4 things.
- The prompt. Ready to paste. Fill in the bracketed parts and send.
- What each line is doing. The line by line read. Take this line out and here’s what breaks.
- Why it works. The short version, if you only read one paragraph.
- One thing, and a simple fix. Where the prompt still falls short, and the sentence that patches it.
Nothing here is a template we made up in the abstract. Every prompt gets run first, and what you get is the version that survived.
2 ways to read it
The newsletter works for 2 types of readers. You don’t have to pick one. Most people shift between them depending on the week.
Read to run. You want prompts you can copy, fill in, and send. Skip to the code blocks. Over time your inbox becomes a searchable library you can pull from whenever you need one.
Read to learn. You want to know why a prompt is built the way it is. That’s the line by line section. You start to notice the same structural moves showing up across different tasks and different models.
The best prompters aren’t the ones who memorize templates. They’re the ones who understand the limitations and can get good results in fewer turns.
Most readers will spend most of their time in read-to-run mode. That’s fine. The learning happens either way.
Where the teardowns come in
The name is still the name for a reason. Some weeks the most useful thing is not another working prompt, it’s watching one fail.
When that happens you get a teardown: the prompt as written, what it produces, the specific structural choice causing it, and the rewrite. Same job as the line by line read, in reverse.
Those issues run when there’s a failure worth the space. The rest of the time you get prompts that already work.
Your inbox is the library
Every issue you receive stays in your inbox. When you need a prompt for a specific task, search your email for a keyword. The picks are short enough that email search usually surfaces the right one.
If you want a more structured library outside email, a snippet manager like SnippetsLab, or even a diary app like Day One is decent. Whatever lets you tag and organize them. Your computer’s default notes app, with tags or folders, works well.
What we don’t cover
Not everything falls within our scope. Here’s what you won’t find in Prompt Teardown:
- Jailbreaks, prompt injection attacks, or anything designed to make models behave in ways their creators didn’t intend
- Image generation prompts (text-to-text and text-to-image prompting are different enough to deserve separate treatment)
- Model rankings or “best model” claims
We don’t rank models. You’ll never see us say “Claude is better than GPT” or vice versa. Different models have different strengths, and the interesting question is always how to adapt your prompting approach to the tool you’re using, not which tool is “best.”