testy.coolPrompt Field Notes
Things I figured out about prompting in specific situations. Not universal rules - each one has a context where it applies.
Writing system prompts or behavioral instructions
Teach it how to think. Give it a mental model instead of a list of "don't do this, don't do that."
Not benchmarked, just vibes - but at scale you hit a ton of outliers you can't predict. A list of prohibitions can't cover what you don't know is coming. A mental model is a catch-all that stays strong without inducing unwanted bias. I'll take that over a blacklist any day.
Your product handles diverse or multilingual inputs
Don't give specific examples - they anchor the model and induce bias.
Say you're processing store names across 30 countries. If your prompt includes "e.g. Walmart, Target" as examples, the model starts pattern-matching toward American retail chains. Instead, describe what you want abstractly: "extract the merchant name as displayed on the page." Let the model generalize - that's what it's good at. Not benchmarked, but every time I've removed specific examples from a multilingual prompt, output diversity got better.
You need the model to parse multi-part instructions clearly
Use markdown headers or XML tags to structure prompts. Not CAPS.
Haven't benchmarked this either - I just don't like CAPS. They feel like yelling and I'm not convinced the model reads them the way we intend. Markdown headers or XML tags give clear, parseable structure. I reach for markdown first, XML when the prompt has nested sections or when I want the model to reference specific blocks by name.
Having an LLM generate descriptions for content that only has a title
Tell it to surface non-obvious information. The title already says what happened - the description's job is to add something the headline doesn't.
When you need the LLM to describe a news item that only has a title, the default is title expansion - rephrasing it longer. Zero value. Dropping "non-obvious" into the prompt pushes it past restatement into why it matters, what shifted, what's counterintuitive. Every time I've added it the output went from "yeah, I knew that" to "oh, interesting."