About
A lab for prompting techniques that hold up.
Hundreds of papers about prompting large language models appear on arXiv every month. Most are narrow, hard to reproduce or impossible to apply without retraining a model. PromptLabster reads them for you and publishes only the ones you can put to work with a prompt.
How a paper gets in
- Collect. Every day we pull new arXiv papers on large language models, prompting, agents and evaluation.
- Score. Each paper gets a lab score from 0 to 100. The score asks four questions: can a practitioner apply it with prompts alone, how strong is the evidence, how large is the effect, and is the idea new.
- Filter. Only papers that score 80 or higher are published here. That is roughly one paper in ten.
- Explain. Each paper becomes a technique page with four views: a full breakdown with the method and ready-to-copy prompts, a plain-language version, a one-screen digest, and a list of roles that can use it.
- Link back. Every page names the original paper and links to it on arXiv, so you can check any claim at the source.
What the lab score means
| Score | What to expect |
|---|---|
| 80–84 | A solid technique with clear evidence. Useful in specific situations. |
| 85–89 | A strong, broadly applicable technique with a large measured effect. |
| 90+ | Rare. Changes how you should prompt for a whole class of tasks. |
How the pages are made
We use large language models to read the papers, draft the explanations and check the output against the source. The scoring rubric, the editorial rules and the page format are ours. Where a paper reports numbers, the page repeats them as published.
What this is not
PromptLabster is not peer review. Each technique reports what its authors measured on their tasks and models. Results on your task may differ, so test a technique on a few real examples before you rely on it.
Questions people ask
Where do the papers come from?
From arXiv, the open archive where most AI research appears first. We check new papers on language models, prompting, agents and evaluation every day.
What does the lab score measure?
How useful the technique is in practice: whether you can apply it with prompts alone, how strong the evidence is, how large the effect is and how new the idea is. We publish papers that score 80 or higher.
Are the pages written by AI?
Language models read the papers and draft the pages. The scoring rubric, the editorial rules and the page format are ours, and every page links to the original paper so you can check it.
Can I trust the numbers on a page?
The numbers are the ones the authors report. They describe the authors' models and tasks, so treat them as a strong hint and test the technique on a few of your own examples.
How often is the site updated?
Every week. About forty new papers a month pass the 80-point bar. The newsletter brings the best of them to your inbox on Sundays.
Do I need an account?
No. Every page is open to read. The newsletter only needs your email address.
How do I unsubscribe?
Every email has an unsubscribe link at the bottom. One click removes you from the list.
Can I suggest a paper or report a mistake?
Yes, please do. Write to [email protected] with the arXiv link or the page and what looks wrong.
Contact
Questions, corrections or a paper we missed: [email protected].