Which tool to choose
- You want an assistant that always replies with your style and your rules: ChatGPT's "custom GPTs" feature or the "projects" with instructions in Claude and Gemini. No code, just written instructions.
- You want it to also perform actions in your apps (save, send, update): connect the custom assistant to a no-code platform like Zapier, which gives it access to the apps.
- You want a truly custom agent, with complex logic: there you need to program, and it's another level; for personal use almost never necessary.
How to do it
- Define the agent's single task. A good personal agent does one thing well: "correct my emails keeping my tone," "turn my notes into action lists." Not "do everything."
- Write the instructions as you would to a new assistant. The operational syntax, to paste into the instructions field:
Role: you are my assistant for work emails.
Behavior: correct grammar and clarity while keeping my direct and cordial tone. Don't lengthen the texts. Don't add courtesy formulas I didn't write myself.
When you're not sure about the tone, ask me instead of guessing.
- Upload the reference materials. Examples of your best emails, a document with your rules, a glossary of your terms: the agent will use them to resemble you.
- Test it on real cases and correct the instructions. The first answers tell you what it misunderstood: you adjust the instructions until the behavior is the right one. The agent is refined by writing, not by programming.
- Add the actions only if needed. If it's enough for it to write, stop here. If you want it to save or send, then you connect an automation platform, with the proper confirmations before the actions that matter.
A concrete example
Luca reads many articles and wants an agent that always summarizes them in the same format. He creates a custom GPT with the instructions: "for every text I give you, return three key points, a verbatim quote, and a question the text leaves open." He uploads three examples of how he wants them. He tests it: the first summary is too long, he adds "maximum thirty words per point." From that moment he pastes an article and gets the card in his format, without repeating the instructions. He built a personal tool in fifteen minutes, writing in plain language.
When it does NOT work (and how to fix it)
If the agent doesn't follow the instructions you gave it
It happens when the instructions are vague or contradict each other. Remedy: make them concrete and with examples ("not like this: [example]; like this: [example]"). An agent follows a rule shown better than one described.
If it forgets the instructions after a while in the conversation
In long chats the initial instructions lose weight. Remedy: in customized assistants the instructions stay fixed and reapply to every new conversation, so open a new chat when you notice it drifting, instead of continuing in the long one.
If you can't find the feature to create a custom agent
The name changes: "GPT," "projects," "Gem," "assistants." Remedy: look in the menu for an item to create a tailored assistant. If your plan doesn't offer it, you get almost the same effect by saving the instructions in a document and pasting them at the start of each conversation.
A tip from someone who really uses it
Keep your agent's instructions in a separate document, outside the platform. It's the "brain" of your tool: if you switch assistants, you rebuild it in five minutes by pasting that document, instead of starting from scratch. The tool is yours only if you own its instructions.
Frequently asked questions
Do I need to know how to program to create a personal agent?
No, for personal uses almost never. A custom assistant is built by writing instructions in plain language and uploading documents. Programming is needed only for agents with complex logic, which for everyday life are rarely necessary.
How different is it from writing a good prompt each time?
The difference is that you write the instructions once and they apply forever, on every conversation, without repeating them. It's a prompt that becomes permanent and to which you can attach your reference materials.
Is it true that a personal agent learns my preferences on its own over time?
It's an expectation that leads to disappointment. Custom assistants follow the instructions you write, they don't autonomously learn your tastes the way a person would. They improve because you update the instructions, not because they observe you: you update the "brain" by hand.