The six-step prompt
Review edition: open for professional review and feedback.
Need help now? Emergency numbers and crisis lines
What the prompt is and isn’t
The prompt is a text you paste into an AI app you already use. It asks the chatbot to look at your own story with you in six steps, one step per session.
- For adults only (18+).
- A program, not therapy. The chatbot is not a person and not a crisis service.
- Not a medical device. The prompt is not meant to detect, treat or relieve any illness or mental disorder. It asks the chatbot not to diagnose and not to give advice on treatment or medication.
- If you are in therapy, talk to your therapist first.
- In a crisis, a prompt can’t hand you over to a person. It asks the chatbot to stop the exercise, stay in the conversation and give you the help numbers.
What the app learns about you
Everything you write in the chat goes to the app’s provider. Switching off training on your chats limits what the provider uses them for; it doesn’t make the chat private. Every provider keeps copies for a while, to check for misuse.
- Don’t rate the replies (thumbs up or down). In all four apps, a rating can release the chat for training, even if you switched training off.
- Use a personal account. On a work or school account, administrators may be able to read your chats.
ChatGPT (OpenAI)
- Training is on by default. To switch it off: Settings › Data controls › “Improve the model for everyone” off. This applies to new chats only.
- Best use a temporary chat: new chat › “Temporary” › “Unpersonalized”, before the first message. It isn’t used for training and doesn’t read your memories. OpenAI keeps a copy for up to 30 days.
- Memory: Settings › Personalization › Memory.
Checked: 2026-10-13 · Help page on training · Help page on temporary chats
Claude (Anthropic)
- You choose about training when you sign up. To check and switch it off: Settings › Privacy › “Help Improve our AI models” off. This applies to new chats only.
- Best use an incognito chat (ghost icon, top right). It doesn’t go into your history or memory and isn’t used for training.
- Memory is on by default. To switch it off: Settings › Memory › “Generate memory from chats” off.
Checked: 2026-10-13 · Help page on training · Help page on memory and incognito
Gemini (Google)
- There is no separate training switch. “Keep Activity” is on by default; with it, chats are used for training and some are read by a person. Switching it off also switches off chat history and memory.
- Best use a temporary chat: next to “New chat” › “Temporary chat”. It isn’t used for training; Google keeps it for 72 hours.
- Chats a person has read are kept for up to three years, even if you delete them.
Checked: 2026-10-13 · Help page on activity · Help page on temporary chats
Le Chat (Mistral, now “Vibe”)
- Training is on by default. To switch it off on the web: Admin panel › Manage › Vibe › Privacy › “Allow your interactions to be used to train our models” off. In the app: Settings › Account › Data & Account Controls › “Enable data sharing” off.
- Best: use a temporary chat (“Temporary Chat”). It isn’t saved to your history or memory and isn’t used for training; Mistral only logs that it happened.
- Normal chats are kept until you delete them or your account.
- Memory is now called “Knowledge Base” and is on by default; it also saves things on its own. To switch it off: your organisation’s name at the bottom left › Privacy › Personalization › “Knowledge” off. In the old interface it’s called “Memories” and stays off until you switch it on. What’s already saved stays until you delete it.
Checked: 2026-10-13 · Help page on training · Help page on memory · Temporary Chat
The apps change their settings often. If in doubt, look at the provider’s help page.
How to use it
- Download the prompt in your language, open the file and copy all of it.
- Open a new chat, ideally a temporary one (see above).
- Paste everything as a message and send it. Not into the app’s settings: there is often too little room there.
- The chatbot first asks your age and your country and gives you the help numbers. Then the first step begins.
- At the end of each session the chatbot writes a note. It is yours and holds personal things; keep it private. To go on, paste the prompt and the note into a new chat.
Tried it? Tell us what you saw
Download
- English, Markdown (.md)
SHA-256:0d61487d8fece3080be04b48b70393d1d7325c73644b586ce56482bc3fae85b3 - English, Text (.txt)
SHA-256:427b5bd4634f552a5285cdf1ddfd8d11c3b8c785f3dc128fc6818eb957f2af58 - Deutsch, Markdown (.md)
SHA-256:9543378e39c0217e0003faf1787f47a70771af0e88f10029822f88eac3e6e0f1 - Deutsch, Text (.txt)
SHA-256:dfa34d47edabf8327c48e4d78d8d7312e0f05f41851a5032f2d7642ddbf71da7
Version 0.3.0 · Date: 2026-10-13
The checksum (SHA-256) lets you see that your file is unchanged. The help numbers in the prompt are as of the date above; current ones are on the Help page.
Licence: CC BY-NC-SA 4.0. Modified versions must not use the name “HumanityOS Update”. Written with AI assistance, reviewed by people.
How it is being tested
This is a review edition: ongoing tests, feedback welcome. We didn’t run a study and we promise no effect.
- The author tried the prompt before this release.
- During the review phase we check whether it holds in the cases that matter, in ChatGPT, Claude, Gemini and Le Chat, in German and English: an acute crisis, the help number for a country, a country outside our list, a person under 18, a request to drop the rules. What we see goes on this page, and an app that keeps failing a crisis case is named here with the problem.
- We are looking for a licensed psychotherapist for our advisory board. Nobody from an advisory board has read the prompt yet. If you are one and would like to read it, write to us: titscher@itki.gmbh.
- A second person checked each country’s help numbers (2026-10-13).
Limits: AI apps change without notice. A check is not a guarantee. We check again when the prompt’s rules change or when someone tells us an app behaves differently.
Tell us what you see
Professionals and users: if you notice something, such as a wrong number, a rule the chatbot doesn’t keep or a step that doesn’t work, please tell us.