How do I get a bot to answer the way I want?
A bot (ChatGPT, Claude, Gemini) behaves the way you set it up at the very start. That setup is called instructions for the bot (often called a system prompt), a short text telling it: who it should be, who it writes for, in what tone, in what format, what it must always do, and what it must not do. Without it the bot guesses, and every chat ends up in a different style.
You fill in the form step by step: who the bot should be (e.g. *"advisor in a shoe store"*), what it should do, who it writes to, what tone (formal, warm, terse, technical), what format (plain text, list, JSON), what it must always do, what it must not do, examples (1-3 question/answer pairs) and what the bot should do when it doesn't know. The generator stitches it into one well-structured instruction that you paste into the bot as the first message.
Plus 4 ready templates for common cases (customer support, code review, content writing, JSON extraction) and a token counter. It shows how much each request will cost. A token is a chunk of a word, the bot splits every text into chunks like that and you pay for each one.
How to use it
- Pick a template close to your case or start blank. Four ready ones cover most common situations.
- Basics: enter who the bot should be (e.g. *"customer support assistant"*), what it should do (e.g. *"answers questions about orders"*) and who it writes to (e.g. *"store customers, mixed ages"*).
- Reply style: pick a tone (formal, warm, terse and to the point, technical, creative, guides through questions) and format (plain text, list, JSON, step-by-step, code only).
- What the bot must always do: concrete instructions. E.g. *"always greet by name"*, *"include the line number in code"*.
- What the bot must not do: e.g. *"do not promise refunds without checking the policy"*, *"do not invent functions that do not exist"*.
- Examples (1-3 question/answer pairs): show the bot exactly what you want back. This works stronger than ten descriptions.
- What the bot should do when it doesn't know: e.g. *"hand off to a human"*, *"return an error in JSON"*, *"say honestly that you don't have the answer"*.
- Copy the final text and paste it into the bot as the first message. The token counter at the top shows how much each request will cost.
When this is useful
Six typical situations where this generator gives you a concrete edge:
- Customer support bot. The *"customer support"* template gives you a ready frame: who the bot is, warm tone, key rules (greet by name, show that you understand the issue, ask one clarifying question, never promise refunds without checking the policy). You tweak it for your company in 5 minutes instead of writing it from scratch.
- Code review assistant. The *"code review"* template sets the bot up to focus on real bugs (not formatting nits), cite the line number, suggest a ready fix, and split *"must fix"* from *"nice to have"*. You take it and run.
- Pulling data into JSON. The *"JSON extraction"* template gives strict rules: return JSON only, nothing else, use *"null"* for missing fields instead of inventing data, return an object with an *"error"* field on failure. Reliable when the bot feeds your system automatically.
- Writing blog content. The *"content writing"* template configures the bot for a sensible style: lead with the problem (not the product), use concrete numbers (not empty adjectives), cut corporate filler. The output reads like a good copywriter wrote it, not like generic AI mush.
- "Try again differently" iterations. Bot replies too long? Switch length to short. Too cold? Switch tone to warm. Unwanted formatting? Switch format to plain text. No rewriting the whole instruction, you change one field, click, copy.
- Teaching the team to write good instructions. Instead of staring at a blank page, a junior gets a form: fill in the fields, the generator assembles a professional instruction. After a few uses they learn to think in sections and write better prompts even without the generator.