Editorial Policy
This page explains how prompts are chosen and written for TokenCraft, how the example run on every prompt page is produced, and what that example does and does not promise.
How prompts are chosen and written
TokenCraft has no user-generated content and no paid placements. Prompts are chosen for common tasks people bring to AI assistants — work, study, writing, code, travel, food — and each is rewritten so that it states a role, the inputs it needs, the steps to follow and the format of the answer. AI tools are used to draft and revise prompt texts; the editor decides what is published.
The English and Russian libraries are two separate collections, not translations of each other: each is written for the tasks and context of its own readers.
The parts of a prompt you are expected to change are turned into variables. Each variable becomes a field on the page — a line of text, a longer text box for pasted material, or a list of choices — and whatever you type is inserted into the prompt before you copy it or open it in a chat.
How the example run is made
Every prompt page shows one real run of that exact prompt text. It is produced in two steps:
- A model fills in the prompt's variables with realistic values. It is instructed to invent people, companies and figures, and to name only places, software and laws that exist; the values are shown, so you can judge them.
- The prompt, with those values inserted, is sent to a model in a fresh session: a one-line system prompt, no custom instructions, no memory, no other context.
The page shows the values that were used, the model that answered and the date of the run. The answer is published as the model gave it: it is converted from Markdown to HTML, its headings are shifted one level to fit the page, unsafe markup is removed, and nothing in its text is rewritten. You can put the same values into the form with one click and see exactly the text that was sent.
Some prompts are conversations: they end by waiting for your reply. For those, the example is the model's opening turn, and the page says so.
An example belongs to the prompt text it was generated from. If a prompt is edited, its old example cannot be published next to the new text; the prompt has to be run again first. This is enforced when the library is built, not left to memory.
Sources and links
Prompts whose answers depend on real sources — legal research, contracts and policies, literature reviews, fact-checking, market and news analysis — are run with web search, and the page says so next to the model. Before such an answer is published, every link in it is fetched, and an answer with a dead link is not published. Links to scholarly articles are also matched by their DOI against the Crossref registry, because many publishers refuse automated requests.
A live link does not make every claim around it correct. Legal and financial answers in particular should be checked against the primary source before you rely on them.
What an example is not
An example is one run, not a guarantee. Models change, and the same prompt can answer differently tomorrow, in another language, or in another model. Model output can be confidently wrong, and nothing on TokenCraft is professional advice — legal, medical, financial or otherwise. Terms of Use.
Corrections
If a prompt or its example is wrong, outdated or misleading, write to hello@tokencraft.ai. Name the prompt and, if you ran it yourself, the model you used. A prompt that no longer works is rewritten or removed; its example is generated again with the new text.