How to use, adapt and check a prompt

Using a prompt from this library

1. Pick by task, not by wording

   Start from what you need to do. Each entry says when to use it and, just as important, when to skip it.

2. Fill in the placeholders

   Words in {{double_braces}} are yours to replace. On an entry page you can type them into the form and watch the prompt update, or copy the prompt and edit it in your tool.

3. Send it, then read critically

   Every entry has a Check the answer list. Go through it before you use the result for anything that matters.

4. Adapt and keep what works

   The How to adapt it notes cover the common changes. Save the prompts you come back to; they stay in this browser.

Placeholders and variables

A placeholder is a slot such as {{audience}}. Each one is listed under Fill in the prompt with a short hint and an example value. The build checks that every placeholder in a prompt is explained and that no explained variable is unused.

Anything you type into the form stays on the page you are reading. It is not saved, not put in the page address and not sent anywhere. It disappears when you reload or leave the page.

Adapting any prompt

Most improvements come from a few changes that apply to every prompt here:

• Say who the result is for. A named reader changes vocabulary, length and what counts as important.
• Ask for the evidence. Quotations, line numbers or a Source column make the answer checkable.
• Allow "I don't know". Tell the model to write MISSING, not set or a question for you rather than guessing.
• Fix the shape of the answer. A table, a numbered list or a fixed set of headings is easier to check than free prose.
• Change one thing at a time. When a prompt stops working, compare the old and new versions on the same input.

To choose between two versions of a prompt, run both on the same inputs and treat the results as options in Compare options against criteria you set in advance, with criteria written before you read the outputs.

What "works with" means

Each entry lists kinds of model rather than brands, because products change faster than the prompts do. A kind marked good fit suits what the prompt asks for; workable means it can work with the limitation stated in the note. Kinds that are not listed were not considered.

General chat assistant

A conversational assistant in a browser or app.

Reasoning model

A model that works through a problem before it answers. Slower, and better at tasks with several steps.

Small local model

A model that runs on your own machine. Your text stays local, but context is shorter and instruction following less reliable.

Model API

A model called from code. Put the fixed rules in the system message and the material in the user message.

Saving prompts

The Save button keeps a shortlist on the Saved page. Only the prompt identifiers are stored, in this browser's local storage under the key prompt-library-saved. The values you type are never stored.

• Saving needs JavaScript. Without it, every prompt is still listed and can be copied by selecting the text.
• The shortlist does not follow you to another browser or device, because there are no accounts.
• Clearing this site's data in your browser removes the shortlist.

What "sample" means here

Every entry in this build is marked as a sample, and the build refuses entries that are not. The prompts are written to be genuinely usable, but they are a small starting set, not an exhaustive or endorsed collection. Version histories and dates are sample records rather than a real editorial log.

Adding or changing a prompt

1. Start a draft

   Run python3 tools/new_prompt.py my-prompt-id "Title of the prompt". It writes content/en/prompts/my-prompt-id.json with the next free code and the state draft.

2. Fill every field

   Run python3 build.py --check. It names every problem with the file and field, including placeholders that are used but not explained.

3. Try it for real

   Use the prompt on realistic input in each model kind you list, and write the Check the answer list from what you saw.

4. Publish

   Set the state to published, add updated_on and a history entry with the same date, then run python3 build.py.

What this library does and does not do