Prompts that hold up in production.
A working library of prompts for Claude, GPT, Gemini and image models, each one built as a specification you can inspect and reuse. Alongside it, essays on why some prompts work and most do not.
Featured prompts
A sample from the library. Every prompt is structured, annotated, and ready to copy.
The prompt library
Filter by category. Click any card to read the full prompt, copy it, and see the reasoning behind its design.
Writing on the craft
Essays on how prompting actually works, past the folklore and the magic phrases.
About Tai Tivalu
Tai Tivalu is a prompt engineer focused on the unglamorous part of the field: making prompts that behave the same way on the thousandth run as they did on the first.
This site is a public notebook. The library collects prompts that have earned their place through real use, each one written as a specification rather than a lucky phrase. The essays work through the principles underneath, so you can build your own prompts instead of borrowing mine.
The through-line is simple. A good prompt is not a clever incantation. It is a clear brief, with a role, real context, a specific task, an explicit format, and honest constraints. Everything here is an argument for that idea, worked out in detail.
- Specification over incantationName the parts of a prompt so you can debug them.
- Information over wordingContext the model lacks beats clever phrasing every time.
- Constraints do the heavy liftingMost bad output is the model doing something you never asked for.
- Test adversariallyA prompt that only works on the happy path is a demo, not a tool.