FAQ

No. Everything happens in your browser. The builder assembles text from templates using ordinary JavaScript — there is no AI model behind it, no server call, and no account. What you type never leaves your device unless you copy it somewhere yourself. That is also why the site works offline once loaded, and why there is no cookie banner.

CRAFT. Five slots — Context, Role, Action, Format, Tone — cover most of what a model needs to stop guessing, and the context-first ordering matches what Anthropic and Google both document. If you work mainly inside Microsoft 365, learn GCSE instead, because it is Microsoft’s own and it has a slot for pointing at your files.

Two of them are. GCSE is Microsoft’s own, documented in Microsoft 365 Copilot support material. TCREI is Google’s, taught in Google Prompting Essentials. CLEAR comes from a peer-reviewed paper. The rest — RTF, TAG, APE, RACE, CRAFT, RISEN, CO-STAR — are practitioner conventions with no vendor endorsement. Every framework page on this site carries a badge saying which it is, because that distinction is usually missing from prompt-framework lists and it matters.

Because the vendors document different preferences. Anthropic recommends XML tags and putting context before the task. Google asks for concise prompts with clear delimiters, and for your instruction to come after a long block of data. OpenAI’s current guidance is to keep prompts lean and avoid repeating instructions. Microsoft Copilot is built for flowing natural language, so heavy markup is wasted there. The builder keeps your content identical and changes only the packaging.

On current frontier models, no. They reason internally by default and the instruction mostly adds tokens and latency; independent testing puts the accuracy gain at marginal while adding 20–80% to response time. Use the provider’s own control instead — effort on Claude, thinking_level on Gemini. On small or open models you host yourself, keep it: Meta still lists chain-of-thought among its official Llama techniques.

The builder itself never transmits anything, and if you save a draft it is stored only in your own browser. But the prompt you build is meant to be pasted into an AI tool, and that tool has its own terms, its own data handling and its own retention. Treat the finished prompt as if you were about to publish it, and check your organisation’s policy before putting client data, personal data or trade secrets into any AI assistant.

Every model and framework page shows a "verified" date, and every claim links to the vendor page it came from. Model lineups change every few weeks, so treat the vendor link as the authority and this site as an orientation layer. If a date on a page looks stale, click through — the vendor page is always right and this one might not be.

Because for most people Copilot is the AI they actually use, and prompting it well is a different skill from prompting an API model. It is grounded in your own tenant, you do not choose the underlying model, and Microsoft publishes its own prompt framework for it. The Microsoft page says plainly that the model underneath is selected by Microsoft and can change.

No — and on current models it can make things worse. OpenAI documents that removing repeated instructions and duplicated examples improves both performance and efficiency, and that repeating rules such as "ask first" causes the model to over-apply them. Google asks for concise prompts outright. Specificity is the variable that helps; length is not.

A fixed list of checks you can read in full next to the score. It looks for things like whether you named an output format, gave the model something to ground against, stated success criteria, and avoided known anti-patterns. It is deterministic and it is not an AI judge — it cannot tell you whether your prompt is right for your task, only whether it contains the elements the vendor guidance says a good prompt contains.

Because they are genuinely different choices. Plenty of Dutch users prefer a Dutch interface while writing English prompts, either because the model performs better in English on their task or because the output goes to an international audience. The two switches in the sidebar are independent.

Yes. "Copy link" puts the whole prompt into the URL itself, so anyone opening it sees your filled-in form and can edit from there. Nothing is stored on a server — the link contains the data. That also means a long prompt makes a long link, and that you should not share a link containing anything confidential.

No. It is an independent reference. Product names and trademarks belong to their respective owners and are used here only to identify the products being described. Nothing on this site is endorsed by, sponsored by, or produced in cooperation with any of them.

Almost always, the instruction was sitting inside a block of pasted content and the model read it as data rather than as a command. Move it out, wrap the pasted material in its own delimiter, and state the instruction once in its own section. If you are working with a long document, put your instruction at the very end — Google recommends exactly this arrangement.