AI Prompt Generator & Prompt Improver

About the AI Prompt Generator & Prompt Improver

A good prompt says who the AI should be, what to do, what it needs to know, who the answer is for, and what the answer should look like. Build a prompt turns a short form into exactly that, on one of four structures — Role–Task–Format, CO-STAR, step by step or few-shot — with sections as Markdown headings, XML tags or plain labels. You get the prompt as one message, a system prompt + user message version for apps and APIs that have a system (developer) message, and a reusable few-shot block that ends in an {{input}} placeholder. The form is checked as you type: a missing role or format, a few-shot prompt without examples, and length limits that disagree.

Improve my prompt takes a prompt you already use and returns a clearer version, a list of what changed and a checklist of what is still missing, with the changes shown word by word. It keeps every {{placeholder}}, number and example of yours; a version that drops them is not shown. It runs on your device where it can, on MySmartCoPilot’s cloud AI only if you choose it, and as a rule-based checklist with no AI at all.

Every prompt gets token counts: exact for OpenAI’s tokenizers, and labelled estimates for Claude and Gemini, whose tokenizers are not published. Prompts that try to get around an AI’s safety rules — jailbreaks, “developer mode”, “ignore your instructions”, “uncensored” or filter-bypass prompts — are refused.

Common uses

• Writing reusable prompts for ChatGPT, Claude, Gemini or any other assistant.
• Turning a one-line request into a prompt that gets the format and length right the first time.
• Setting up a system prompt and user message for an app, a bot or an API call.
• Teaching a model a format or a classification with a few examples.

The four structures

• Role–Task–Format: who the AI is, what to do, then context, audience, format, tone and rules. The all-round choice.
• CO-STAR: Context, Objective, Style, Tone, Audience and Response — for writing that has to land with a particular reader.
• Step by step: numbered steps the AI works through (yours, or four general ones) before it answers.
• Few-shot: your examples of input and output, in one consistent format, then the new input.

The parts follow the vendors’ own guides: OpenAI’s prompt engineering guide (identity, instructions, examples and context, with Markdown or XML to mark the sections), Anthropic’s prompting best practices (a role, context, examples and XML tags) and Google’s prompt design strategies (clear instructions, few-shot examples in a consistent format, context and clear delimiters).

What the checks look for

• Task missing: the one thing every prompt needs.
• Role, output format or length missing: the AI has to guess them.
• Length limits that disagree — “about 200 words” in Length and “under 100 words” in Rules — are named, so you keep one.
• A few-shot prompt without examples, or an example without its output.
• In the improver: vague words such as “not too long” or “make it good”, and each part the prompt lacks.

An AI-improved prompt must keep at least 60% of your prompt’s key words and every {{placeholder}} and number, or it is not shown.

Token counts

Counts for OpenAI models are exact: the page runs OpenAI’s published tokenizers (o200k_base for the GPT-4o, GPT-4.1 and GPT-5 families, cl100k_base for GPT-4 and GPT-3.5 Turbo), as the LLM Token Counter does. Claude and Gemini do not publish their tokenizers, so their counts are estimates from the providers’ own rules of thumb — about 3.5 characters a token for Claude (about 30% more tokens on the newer tokenizer) and about 4 for Gemini — and are labelled as estimates.

Limitations

• A better prompt makes a better answer more likely, not certain: test it in the AI you use, and adjust.
• The page does not run your prompt in ChatGPT, Claude or Gemini, and there is no test run here.
• The rule-based checks read English prompts; the AI improver works in the language of your prompt.
• Claude and Gemini token counts are estimates. On-device models need a browser with WebGPU; the cloud AI has a small shared daily allowance that resets at 05:30 IST (00:00 UTC).
• The improver takes prompts of up to 2,000 characters (about 300 words).

Privacy

The prompt builder, its checks and the token estimates work in your browser; the exact OpenAI token counts download the tokenizer from MySmartCoPilot once. Improve my prompt runs on your device where it can: in your browser’s built-in AI, or in an AI model that is downloaded from MySmartCoPilot once (its size is shown first). Only if you choose MySmartCoPilot’s cloud AI is your prompt sent, with email addresses, phone numbers and ID numbers masked first; MySmartCoPilot doesn’t store it. If the consent step offers you an outside AI service with more answers a day and you choose it, the company that runs it may keep what is sent and use it to train its AI. This page never sends your prompt to the AI assistant you write it for: you paste it there yourself.