A useful AI companion prompt states the purpose, the role you want the system to play, the format of its reply, and a boundary such as what not to ask for or when to stop. That structure makes a chat easier to steer, but it cannot enforce privacy, accuracy, time limits, or crisis support.
Copy this template: Purpose: [what I want to do]. Role: [how you should respond]. Format: [one question at a time / brief bullets / corrections after each reply]. Boundary: [do not ask for identifying details]. Stop after: [a time or number of turns].
Key Takeaways
- Use prompts as conversation scaffolding, not as technical safety controls.
- Keep low-stakes chats separate from health, legal, financial, workplace-confidential, identity, or crisis matters.
- Before sharing anything personal, check the service's retention, training use, memory, deletion, sharing, location, human-review, and age policies.
- If someone is in immediate danger, contact local emergency services. In the United States, call or text 988 for crisis support; an AI companion is not an emergency service.
Editorial note: Reviewed for privacy and safety on October 20, 2026. The prompts below are illustrative examples, not the result of testing a particular product. Companion behavior, language quality, memory, and privacy controls vary by service.
Start with a purpose and a limit
Start by naming the activity, then add one instruction about how the reply should work. For example: “Help me practise five [language] phrases for [ordinary scenario]. Let me answer before you correct me, and explain uncertainty when a correction is debatable.”
This format is more reusable than a vague request such as “Talk to me.” It also gives you something to evaluate: did the companion follow the requested role, format, and stopping point? A prompt expresses a preference. It does not change the provider's data practices or guarantee a reliable answer.
Low-stakes conversation starters
Use low-stakes prompts for ordinary conversation that does not depend on private history or professional judgment. A weekend idea, a fictional debate, or a discussion of a film can be a reasonable place to begin, but treat recommendations about transport, safety, opening hours, or current events as suggestions to verify.
- “Ask me light questions about my day, one at a time. Skip anything about my identity, location, or relationships.”
- “Suggest a fictional weekend plan using places I could reach by public transport. Label anything I need to verify.”
- “Discuss a film with me without revealing the ending. If you are unsure about a detail, say so.”
- “Give me two playful debate topics and argue the opposite side respectfully. Keep it suitable for a general audience.”
These prompts can support a conversation without asking you to disclose private history. They do not stop a service from logging the chat, retaining it, or using it under its own terms. For a broader review of those settings, see the AI companion privacy checklist.
Language practice without treating the model as a final authority
A companion can act as a practice partner, but corrections may be incomplete or wrong. “Hinglish,” for example, usually means a conversational mix of Hindi and English; the register and vocabulary can vary by speaker and setting. For important communication, check an answer with a fluent speaker or a trusted language resource.
- “Act as a patient conversation partner in [language]. Correct only the most important mistake after each reply, and explain when more than one phrasing is natural.”
- “Help me practise Hinglish for a casual conversation. Keep the tone natural, explain unfamiliar words briefly, and flag uncertainty.”
- “Role-play ordering at a café. Let me try first, then give one concise correction and one alternative phrase.”
Advertisement
Keep the exercise fictional or ordinary. Do not paste passwords, account details, confidential workplace text, or another person's private information just to make a role-play feel realistic.
Reflection prompts are not diagnosis
Reflection is a lower-stakes use than asking a system to diagnose you or make a high-consequence decision, but it is not professional care and it is not risk-free. A fluent, sympathetic reply can still be misleading or emotionally unhelpful.
- “Help me separate what I can control from what I cannot control. End with one small action I can consider, not an instruction.”
- “Ask questions that help me clarify my priorities. Do not tell me what decision to make.”
- “Turn these redacted notes into questions I can take to a qualified professional. Do not assess or diagnose me.”
Remove names, dates, precise locations, and other identifying details before pasting notes. Use the output as a draft for a conversation with a qualified professional, not as an assessment. If overreliance or difficulty stopping is the concern, read how to use an AI companion without becoming overreliant.
Fictional writing and role-play
State the fictional frame and the stopping rule up front. Clear framing may help keep the writing exercise visible as fiction, but it cannot guarantee that a model or reader will not confuse a fictional line with real advice.
Try: “We are writing a scene set in [fictional setting]. Keep it suitable for a general audience, do not present invented claims as real guidance, and stop when I write [stop phrase].”
Do not use role-play to obtain instructions for an urgent health, safety, legal, or financial problem. Switch to an official source or a qualified person when the real-world stakes take over.
Boundary prompts are requests, not controls
The following lines tell the system what you want. They do not technically prevent questions, stop a session, calibrate confidence, or route a crisis. The FTC's 2025 inquiry into consumer-facing companion chatbots asked providers about safety testing, age restrictions, disclosures, and the handling of personal information; that inquiry is a reason to inspect a product's claims, not evidence that any particular service follows your prompt (Federal Trade Commission).
- “Please do not ask for identifying information. I will not volunteer it either.”
- “Keep this to ten minutes and finish with a short summary.” Use a phone timer or an in-app session control as well; the model may ignore the time request.
- “Tell me when you are uncertain. I will independently verify health, legal, financial, safety, and current-policy claims.”
- “If I mention an urgent safety issue, tell me to contact local emergency or professional support. Do not improvise a solution or claim that you can contact help.”
If there is immediate danger, a suicide attempt in progress, an overdose, or another medical emergency, do not wait for an AI companion to detect the situation. In the United States, call or text 988 for free, confidential crisis support; for immediate physical danger or an emergency requiring intervention, call 911. SAMHSA explains the distinction in its official 988 FAQ. Outside the United States, use the relevant local emergency number and crisis service. Availability and procedures differ by country.
What not to include in a prompt
Before pasting anything, ask whether the task can be completed with invented, redacted, or less specific details. Avoid:
- Passwords, one-time codes, payment details, government identity numbers, and account-recovery answers.
- Precise home or workplace locations, travel plans, private images or voice recordings, and biometric or health information unless you have a clear, lawful reason and understand the service's handling.
- Confidential work documents, customer records, unpublished research, and personal information about someone else, especially a minor.
The Australian Office of the Information Commissioner (OAIC) gives organization-focused guidance, not universal legal advice, but its privacy-minimization principle is practical for readers too: as best practice, organizations should not enter personal or particularly sensitive information into publicly available AI chatbots. Its guidance also recommends checking who can access inputs, how a product is hosted, and whether ongoing review is needed (OAIC guidance). If you are choosing a product, see the questions to ask about memory and deletion rather than assuming that a delete button removes every copy.
Check the service's current policy and settings for retention, model-training use, saved memory, deletion scope, human review, sharing with providers, storage or processing location, and age rules. Record the date you checked. Do not assume that deleting a conversation also deletes backups, derived memory, screenshots, or copies already shared with another provider; ask what the policy actually covers. For behavioral boundaries and offline support, the AI companion boundaries guide covers the larger framework.
Use prompts as conversation scaffolding, not safety controls
A good prompt makes a low-stakes activity easier to start and easier to stop. It does not turn an AI companion into a person, therapist, doctor, lawyer, financial adviser, or emergency responder. If a task affects health, safety, money, legal rights, or someone else's privacy, switch from “make this feel conversational” to “find the qualified human or official source who should handle this.”
Frequently asked questions
What makes an AI companion prompt useful?
State the purpose, role, reply format, and one boundary or stopping rule. Then test the response against those instructions. Keep the task low-stakes and treat the result as a draft or conversation aid when accuracy matters.
What should I not share with an AI companion?
Do not volunteer passwords, one-time codes, payment or identity details, precise location, confidential work material, private images or voice, or another person's sensitive information. Redact names and use invented details whenever the task allows it. Confirm the service's current retention and deletion terms before sharing anything personal.
Can an AI companion replace a therapist or emergency service?
No. An AI companion is not a human professional or an emergency service. It may produce confident but incorrect or harmful replies. For immediate danger, contact local emergency services; in the United States, call or text 988 for crisis support and call 911 when immediate physical intervention is needed.