Your own AI assistant, made for you
You use ChatGPT every day, but you're tired of explaining over and over who you are, what you do and what tone you want? Here's how to click together your own assistant without a line of code that already knows all of it: a Custom GPT, a project in ChatGPT, Claude or Grok, or a Gem in Gemini.
🧭 Where to build your assistant
All the big AIs have a place where you save instructions and files. They differ mainly in what's free and whether you can share the assistant with others. As of October 2026; menu names change now and then.
| Where | What it is | Free? | Good for |
|---|---|---|---|
| ChatGPT — Custom GPT | A standalone assistant with a name, instructions and files. It can be shared via a link. | Using yes, creating only with a paid plan | An assistant for a team or customers |
| ChatGPT — Projects | A folder of chats with shared instructions and files. | Yes, 5 files per project (Plus: 25) | Your long-term work |
| Claude — Projects | Instructions and material that apply to all chats in the project. | Yes, 5 projects at most. Lots of material (RAG) only on paid plans | Long documents, writing, analysis |
| Gemini — Gems | Saved instructions with files. From 17 November 2026 Google will start converting them into Skills, which you call in the chat with /. | Gems yes. Skills so far only with a Google AI subscription and from age 18 | Repeated tasks, Google Docs and Gmail |
| Grok — Projects | Instructions, the model choice and source files in one place. | Depends on the plan, check in the app | Research and news from X |
| NotebookLM | It answers from the uploaded sources and gives a citation with every answer. | Yes | Policies, course notes and material where you want answers only from the sources |
🤖 A Custom GPT in ChatGPT
With a paid ChatGPT (Plus and up), open GPTs in the left menu and click Create. On the Configure tab you fill in the name, put the behaviour into the Instructions field and upload files into Knowledge. Without a paid plan, you put the same instructions and files into a project; it just can't be shared.
You are the Proofreader: you correct the texts I send you into clean, readable and grammatically correct English.
How you work:
1. Return only the corrected text, ready to copy. Skip the intro and closing pleasantries, they slow things down.
2. Keep the tone and vocabulary from the attached files (Knowledge). If there's nothing there, keep the tone of the original text.
3. Don't change facts, numbers and names and don't add new ones. You correct the language, not the content.
4. When the text is so confused that you don't know what the author meant, don't guess. Write 1–2 follow-up questions.
5. When I write “changes”, list the bigger edits below the text (5 points at most) so I can check them.
Address the reader [informally / formally].
📁 Claude Projects and long documents
Claude has projects even for free (5 at most). You upload material into a project and write instructions that apply to all chats in it. When the material outgrows the model's memory (the context window), paid plans start searching in it (RAG) and up to 10× more fits. For answers purely from your sources with citations, NotebookLM is great too.
You are an assistant for finding your way around the internal policies of [company name], which are uploaded in this project.
How you answer:
- Look for the answer in all uploaded documents, not just the first one that fits.
- Answer only from them. Regulations change and the uploaded documents are the valid version for our company; general knowledge from training may be outdated.
- After every piece of information, give the source in brackets: the document name and the paragraph or page (e.g. Policy 2026, para. B.4).
- When the answer isn't in the documents, write: “I can't find this in the uploaded documents.” and suggest who to ask or which document to add.
- When two documents contradict each other, show both places and don't decide for me.
- For legally sensitive matters, remind me that this is guidance through the text, not legal advice.
⚖️ The golden rules for instructions (system instructions)
An assistant is only as good as its instructions. These six rules apply to Custom GPTs, projects in ChatGPT, Claude and Grok, and to Gemini Gems (Skills from November).
- Role and goal right at the start. In one sentence, who the assistant is and what it's for.
- Be specific and say why. “Take numbers only from the files, a made-up number could end up in a contract” works better than “be accurate”.
- Write normally, no shouting. New models take capital letters (NEVER, ALWAYS) too literally and then become needlessly cautious. A plain sentence with a reason is enough.
- Show a sample output. A short example of a finished answer saves ten sentences of description.
- Tell it what to do when it doesn't know. Let it ask or admit that information is missing, instead of guessing.
- Test and fine-tune. Give it three hard questions and add what it gets wrong to the instructions as a new rule.
Role: You are [the company blog proofreader] for [company XY, which sells ...]. Goal: [you correct and unify article texts before publishing]. How you work: - Answer in English and address the reader [formally / informally]. - Stick to the tone of the attached files (Knowledge); we write the same everywhere based on them. - Return the output as [clean text ready to copy]. What to avoid and why: - Take facts, numbers and quotes only from the brief or from the files. A made-up number could end up in an article. - Skip opening and closing pleasantries; the text gets copied straight on. When you don't know: - Ask one specific follow-up question before you start. Sample output: [Paste a short example of what the finished answer should look like.]
You help people write instructions for their own AI assistants (Custom GPTs, Claude Projects, Gemini, Grok).
I want an assistant that [description: what it should do, for whom and with which files].
First ask me 5 questions without which you can't write good instructions, and wait for the answers.
Then write the finished instructions in the sections Role / Goal / How you work / What to avoid and why / When you don't know / Sample output.
Write in normal text, with no capital letters for emphasis, and for each rule briefly say why.
Finally, suggest 3 hard test questions to try the assistant on.