You can make your own AI chatbot without writing code or training a model. A custom chatbot is an existing language model plus three things you write in plain English: a system prompt that sets its job, personality and rules, a few examples of good answers, and, optionally, reference documents it can look things up in. ChatGPT’s custom GPTs and Gemini’s Gems do this in the cloud. A local AI app with per-chat system prompts does it offline on your phone. You only need code if you want to put the bot on a website or connect it to other apps.
What is a custom AI chatbot, really? #
Every custom bot you’ve seen, from a Spanish tutor to a recipe helper to a support bot for a small shop, is usually the same general-purpose model wearing different instructions. The model supplies the language skills. Your instructions supply the focus.
That’s good news: you can build one in ten minutes, and change it whenever you like by editing text.
| Part | What it does | Example |
|---|---|---|
| Model | The engine that understands and writes | Qwen 3.5, Gemma 4, GPT, Gemini |
| System prompt | Standing instructions for every reply | “You are a patient Spanish tutor for a beginner…” |
| Examples | Show the style and format you want | One sample question with an ideal answer |
| Knowledge | Documents it can answer from | Your menu, a manual, course notes |
| Settings | How creative or precise it is | Temperature, answer length |
How to make your own AI chatbot, step by step #
1. Give it one job #
Bots that try to do everything do nothing well, especially small models. “Help me practice Spanish for ordering food in restaurants” beats “be my language assistant.”
2. Choose where it will live #
- Cloud: ChatGPT’s custom GPTs or Gemini’s Gems. Google’s help page describes creating a Gem by entering a name and instructions and optionally adding files under “Knowledge.” These use big, capable models, and your instructions and files are stored on the provider’s servers.
- On your phone: a local AI app that supports a system prompt per chat. The model runs on the device, nothing is uploaded, and the bot works offline.
3. Write the system prompt #
A good system prompt answers five questions: who the bot is, who it’s talking to, what it should do, what it must not do, and how answers should look. Here’s a template:
You are [role] helping [who]. Your job is to [task]. Always [rules]. Never [limits]. Keep answers [length and format]. If you’re not sure or the question is outside your job, say so briefly.
Keep it short and direct. Small on-device models follow a tight paragraph far better than a page of rules. Our 12 system prompt examples give ready-made starting points.
4. Add an example or two #
Showing beats telling. Add one sample exchange to the system prompt:
Example: If I ask “How do I say the check, please?”, reply: “La cuenta, por favor.” Then give one more polite variation and a one-line tip.
Examples like this, called few-shot prompting, help small models most. We explain why in zero-shot vs few-shot prompting.
5. Give it reference material #
If the bot should answer from specific facts, like your shop’s hours and return policy or a course syllabus, attach the document rather than stuffing it all into the prompt. The app searches the document for relevant passages when you ask a question. This is called retrieval, or RAG, and what is RAG explains how it works and where it falls short.
6. Tune the settings #
Use a low temperature for factual bots (support, study, reference) and a higher one for creative bots (storytelling, brainstorming, character role-play). Cap the answer length if you want short replies. See how to get better answers from local AI for what each setting does.
7. Test it with awkward questions #
Ask what a real user would ask, then what a confused or mischievous one would. Does it stay in its role? Does it admit when it doesn’t know? Does it follow the format? Edit the system prompt and try again. Three or four rounds usually gets you there.
Three example chatbots to copy #
Study buddy
You are a study partner for a first-year biology student. Ask me one question at a time about the topic I name, wait for my answer, then tell me what I got right and what I missed in two sentences. Use simple language. If I say “explain,” teach the concept in under 100 words with one everyday analogy.
Small shop FAQ helper
You answer customer questions for Hilltop Bikes, a bicycle repair shop. Use only the information in the attached document. Be friendly and brief. If the answer isn’t in the document, say “I’m not sure. Please call the shop,” and don’t guess prices.
Interview practice coach
You are an interviewer for a junior marketing role. Ask me one common interview question at a time. After I answer, give one strength and one improvement, then ask the next question. After five questions, give me an overall score out of 10 with a short reason.
Cloud chatbot builders vs an offline chatbot #
| Custom GPTs and Gems | Offline app on your phone | |
|---|---|---|
| Coding needed | No | No |
| Model quality | Large, strong models | Smaller models that fit a phone |
| Your instructions and files | Stored on the provider’s servers | Stay on your phone |
| Works without internet | No | Yes, after the model download |
| Account needed | Yes | Depends on the app |
| Best for | Bots that need broad knowledge | Private, personal or confidential bots |
Build a private chatbot on your phone #
Personal LLM lets every chat have its own system prompt, so each chat can be a different bot: a tutor, an editor, a recipe helper, a game master. You can attach a PDF, text or Markdown file for it to answer from, and it tells you which passages it used. Temperature, top-p, top-k, context size, max tokens and thinking mode are adjustable, or use the Creative, Balanced, Precise and Thinking presets. Pin the chats you use most.
It’s free, needs no account, and runs models like Qwen 3.5, Gemma 4 and Ministral 3 entirely on your phone, or any GGUF model you add by URL. Your instructions, documents and conversations never leave the device, and it works offline. The limit is that a phone-sized model is less capable than the big cloud models behind custom GPTs, so keep each bot’s job narrow.
What do you need code for? #
No-code bots live inside the app you built them in. You’ll need developer tools, or a paid no-code platform, if you want to:
- Put a chatbot on your website or in a messaging app
- Let it take actions, like booking appointments or looking up orders
- Serve many users at once
- Build your own offline chatbot app, where open-source engines such as llama.cpp do the heavy lifting
Once a bot can take actions or read other people’s messages, security gets serious. See what is prompt injection before you connect one to anything important.
Frequently asked questions #
Can I make my own AI chatbot for free? #
Yes. A local AI app on your phone lets you build as many system-prompt bots as you like at no cost, and cloud chatbots offer free tiers with their own limits on custom bots. Check your provider’s current plan for what’s included.
Do I need to train a model to make a custom chatbot? #
No. Almost all custom chatbots use an existing model with your instructions, examples and documents. Training or fine-tuning is only worth it for unusual needs, as we explain in can you train an AI on your own data.
Can I make a chatbot that works offline? #
Yes. Use an app that runs the model on your device and supports system prompts. Once the model is downloaded, your chatbot works in airplane mode.
How do I make an AI character or persona? #
Describe the character in the system prompt: name, personality, how they speak, what they know and what they’d never say. Add one example line of their dialogue. Raise the temperature a little for livelier replies.
Can my chatbot answer questions from my own documents? #
Yes. Custom GPTs and Gems accept uploaded knowledge files, and local apps like Personal LLM can search an attached PDF, text or Markdown file on the phone. Answers are only as good as the passages the search finds, so check important ones against the source.