How every prompt is made and tested.

1. Writtento the rules below
2. Scoredon the ten checks
3. Comparedwith every prompt already in
4. Numberedonce, for good

The rules for every prompt

• Every detail only you know is a blank. Names, numbers, dates and facts appear as blanks like {{role_title}} for you to fill, never as made-up companies or figures.
• Each blank says what goes in it, in one line, and blanks for pasted material such as notes or a document are marked as long, so they open as a large field.
• An example only shows the form. Where a prompt includes an example, it says it is one, and it shows the shape of the answer wanted rather than facts to reuse.
• Each prompt says why it works, in two or three notes, each tied by name to one of the ten checks.
• Prompts for apps and assistants keep the end user's message in one marked place, so the Chatbot Safety Test can try its attacks exactly there.

The score it needs before it is published

Every prompt is read by the same ten checks as the free Prompt Grader, with its system prompt and its prompt taken together, for the model it is laid out for. The repository accepts a prompt only at 9 or 10 out of 10. Today 1,017 prompts score 10, and the 7 that score 9 are written for image and video models, where some of the checks read differently.

The ten checks ask whether a prompt says what to make, who it is for, the background facts, the length, the format, what to avoid, what good looks like, an example, what to do if unsure, and whether its layout suits the model. The ten checks, explained.

A new prompt is also compared with every prompt already in the repository. One whose title is close to an existing prompt is flagged, and goes in only if it does a different job.

Once in, a prompt gets a registry number, such as PPL-001, the first prompt in the hiring collection. The number never changes and is never given to another prompt, so a link or a citation keeps pointing at the same prompt.

Laid out for the model that reads it

Models read prompts differently, so each prompt is laid out for one of eight kinds: Claude, GPT, Gemini, any chat model, a coding agent, Midjourney, an image model that takes plain language, or a video model. The layout is part of the score, and the repository shows which kind each prompt is for. Every chat prompt also works in the other chat models.

Numbers, prices and our own accuracy

• No invented numbers. A figure on this site comes from a source we link, or from a measurement we describe with its date.
• Prices carry their date. AI model prices are checked against each provider's own pricing page on the first day of every month, and every page that quotes them says when (the price table).
• Our tools are measured on test sets kept off the site, labeled by hand before any rule is tuned to them.
• Rules are tuned on one part of each set and tested on another they have never seen, and we publish both results with their dates (the latest figures).

How we write

Plain American English, for someone who has never heard the words prompt engineering. Dates are written out, like October 20, 2026. We leave out hype and the stock phrases that make writing sound machine-made, the same ones the Robot-Speak Checker marks in your own drafts. A score is a count of checks passed, never a mood, and an example is labeled as an example wherever it appears.

When something is wrong

If a prompt gives a poor answer, or a page says something that is out of date or not true, write to contact@thinkhard.si with the page's address. The person who builds the site reads every message. What changes on the site is listed in the changelog.