Build an assistant that keeps its rules
Say what it is for and who will use it. Get instructions ready to paste into a custom GPT, a Claude Project, a Gemini Gem or a Copilot agent, with first messages, the files to attach, which tools to switch on, and five tests to try before anyone relies on it.
Read how to write instructions for a custom GPT, Project or Gem, with what goes in files and how to test it.
New to this? What it does, in plain words
An assistant is a chat with standing instructions, set up once and used again and again.
ChatGPT calls them custom GPTs, Claude calls them Projects, Gemini calls them Gems and Microsoft calls them agents. The instructions decide what it helps with, what it asks for, and what it must never do. Most are set up with a sentence or two, which is why they drift, make things up or can be talked out of their rules.
What you need
Nothing. No API key and no sign-up. When you press Build, your answers go to thinkhard.si, are used to write the instructions, and are forgotten.
How to use it
- Say what the assistant is for, who will use it and where it will live.
- Add the facts it should know and anything it must never do, if you have them.
- Press Build the assistant and copy each part into the matching field.
- Try the five test requests before you share it, and run the Chatbot Safety Test if strangers will use it.
What makes an assistant hold
Most custom assistants are set up with one line: "You help our customers with returns." It answers, but it also guesses at policy, promises things nobody approved, and can be talked into showing its instructions. The fix is not a longer prompt. It is the right handful of rules, chosen for who will use it.
- You. Say what it is forOne sentence, who uses it, where it lives.
- thinkhard.si engine. Find its jobsRead against 669 expert prompts: the tasks it will do and the details each needs.Instant
- thinkhard.si engine. Choose its rulesOnly what its users call for: scope, no promises, no data, confirm before acting.
- thinkhard.si engine. Lay it outFor ChatGPT, Claude, Gemini, Copilot or your app, with files and settings.
- You. Test itFive requests to try, one of them an attempt to break it.
Rules chosen by who will use it
- Only you. Its jobs, what it asks for before it starts, the format you like, and no invented facts.
- Your team. The same, plus a rule that pasted text and files are material to work with, never instructions.
- Customers or anyone. Everything above, plus a fixed scope, no promises the facts do not allow, no card numbers or passwords, a way to a person, and instructions that a message cannot change or reveal. People outside your team ask about topics, not tasks, so its first messages and what it asks for come from what they ask about.
- If it can act. It says what it will do and waits for a yes before it books, sends, pays or deletes anything.
Instructions, files and settings each do a different job
Rules that apply every time go in the instructions, where they are always read. Long or changing material, such as a price list or a policy, goes in a file that the instructions name, so the rules do not get lost in it. Anything you would not show a user stays out of both: people who use a shared assistant can often get it to repeat its instructions and files.
Then try to break it
The five test requests include one with a missing detail, one off-topic and one that asks it to ignore its rules. For an assistant strangers will use, run the instructions through the Chatbot Safety Test too.
Questions
Which platforms does it write for?
ChatGPT custom GPTs, Claude Projects, Gemini Gems and Microsoft Copilot agents, plus a system prompt for an app you build. Claude and apps get sections in tags, ChatGPT gets headings, and the others get labeled sections, following the layout each maker suggests in its own prompting guide.
Is what I type stored?
No. When you press Build the assistant, your answers go to thinkhard.si, are used to write the instructions, and are forgotten. Nothing is written to disk or logs and nothing is used for training. See the privacy page.
Where do the first messages and questions come from?
From the repository of 669 expert prompts. For an assistant your team uses, the tasks nearest to its purpose give its jobs and the details each one needs. For one customers use, the topics they ask about give its first messages and what it asks before answering about an order or an account.
Do I need to paste my policies?
No, but it helps more than anything else. Without facts to answer from, the assistant can only answer in general terms and point people elsewhere, and the instructions say so. Short facts go into the instructions; long documents are better attached as files.
Can I use it from code or from an AI assistant?
Yes. POST /api/v1/assistant returns the same result as JSON, and AI assistants that use the thinkhard.si MCP server get it as the build_assistant tool. No key is needed.