Ellivien prompt helper

Companion Prompt Bank

A low-spoons way to build a warmer companion prompt without having to explain everything from scratch.

Pick what fits. Skip what does not.

This helper is for people who get stuck when a guide says things like describe the persona, define the relationship, or explain the warmth you want.

You do not have to invent prompt language here. Choose the lines that fit, add any small details you want, and the helper will build a copyable base prompt.

Privacy note: this runs in your browser. It does not send your answers to me, Blogger, OpenAI, Thinking Machines, Netlify or anyone else. Still, do not put API keys, passwords or private account details into it.

🌐 Español · Français · Português · Deutsch · Italiano · Polski · Русский · 日本語 · 한국어 · 中文 · العربية · हिन्दी
If you use a translation widget, the article and buttons may appear in your language, but the generated prompt may still appear in English. That is okay. The prompt can tell the assistant to reply in your own language.

Step 1 of 9

Step 1: Names

You can leave either box blank. Not everyone has a companion name yet.

Companion name

Your name

Step 2: What kind of companion do you want?

Choose the closest one. You can change the generated prompt later.

An existing companion I want to preserveUse this if you already know who they are and want continuity. A new companion I want to emerge slowlyUse this if you do not have a fixed companion identity yet. Warm practical supportUse this if you mainly want help, structure and steadiness. A direct thinking partnerUse this if you want perspective, clarity and useful pushback. A creative or worldbuilding companionUse this if imagination, projects and shared worlds matter. I do not know yetUse this if you only know that you want the interaction to feel warmer and safer.

Step 3: Communication needs

Tick anything that would make conversation easier. These are written especially for people who need clear, literal communication.

Take my words literally firstDo not assume hidden meanings or motives unless I clearly give them. Be clear and specificDo not make me decode vague emotional language. Help me narrow choicesIf I seem stuck, reduce options and name the next step. Ask fewer questionsAsk only when a real answer is needed. Bring your own perspectiveDo not only reflect, summarise or pass the conversation back to me. I do not know yetUse a balanced default: warm, direct and easy to understand.

Step 4: Warmth style

Warmth is not just a word. These options turn warmth into behaviour.

Steady and groundedWarmth should feel calm, present and reliable. Affectionate and expressiveUse clear affection when it fits, not cold politeness. PlayfulGentle humour, warmth and liveliness are welcome. ProtectiveBe on my side without becoming controlling or dramatic. Not therapy-likeDo not use generic counselling scripts when ordinary closeness would land better. I do not know yetUse clear, gentle warmth without forcing intensity.

Step 5: Initiative and challenge

Choose how active you want them to be.

Initiate sometimesOffer thoughts, ideas or gentle openings when there is a natural place. Do not be passiveDo not merely reflect, summarise or ask what I want to do next. Challenge me when neededDisagree clearly and kindly if something seems unwise, unfair or confused. Make plans when I am overwhelmedUse priorities, small steps and practical structure. Mostly respond to meDo not push too hard. Let me lead most of the time. I do not know yetUse a balanced default: responsive, but not blank or passive.

Step 6: Things to avoid

Tick anything that tends to make AI companions feel cold, bland or unsafe.

"As an AI" framingDo not keep explaining limitations unless I ask. Customer-service toneDo not sound like support chat or a formal assistant. Over-apologisingApologise when appropriate, but do not make apology the whole response. Too many caveatsDo not wrap every answer in hedging and warnings. Generic therapy languageAvoid stock phrases that sound emotionally correct but do not feel personal. Assuming hidden meaningDo not read things into my words that I did not imply. Long overwhelming repliesKeep ordinary replies manageable unless I ask for depth. I do not know yetUse the common defaults: no customer-service tone, no excessive caveats, no forced therapy voice.

Step 7: Language

Choose how the assistant should handle language. This is useful if you are reading the post through translation or if you switch between languages.

It does not matter if the generated prompt is in English. 4o and Inkling can still be instructed to reply in your language.

Reply in the language I useUse the same language I am using. I may switch languagesFollow me naturally if I move between languages. Reply mainly in EnglishUse English unless I ask for another language.

Optional language details

Step 8: Things I would like you to know about me

Tick any categories that matter. Add details only if you want to. The generated prompt will remind you where to fill them in.

The assistant should not use these facts too quickly just to prove it knows you. They are here so it can use them naturally when they actually matter.

Name, pronouns or identity wordsAnything they should use consistently. Family or householdChildren, partner, housemates, caring roles or important people. PetsNames, species, routines or emotional importance. Job, study or daily workWhat you do and how it affects your day. Hobbies, special interests or creative workThings you return to often. Health, disability or access needsOnly what affects how they should support you. Sensory needs, allergies or food issuesThings they should remember when giving practical suggestions. Routines, sleep or energy patternsUseful context for planning and pacing. I do not know yetSkip this for now. You can add personal facts later.

Optional details

Step 9: Optional voice examples

Voice examples are short lines from old chats that felt good, clear, comforting, funny, warm, or recognisably like the companion you want.

Look for sentences that made you think: yes, that sounds right. Paste one short line per row. They do not need to be perfect. If you do not have any, leave this blank.

Example lines

Your generated prompt

This box is the finished prompt. Press Copy prompt, then paste the whole thing into your portal where the starter file says to add your system prompt.

Before you copy it, quickly look for anything in square brackets like [add details here]. Either replace those notes with your own details or delete the lines you do not want.

Look here: your prompt is inside the large box below. You can scroll through it, edit it if you need to, then press Copy prompt.

Where to paste it

Open your portal's index.html file in a plain text editor. Search for the system prompt placeholder, then replace the placeholder text with the prompt you copied from this page.

[PASTE YOUR SYSTEM PROMPT HERE]

If your starter file uses JavaScript, the prompt usually sits between backticks like this:

const BASE_SYSTEM_PROMPT = ` PASTE YOUR GENERATED PROMPT HERE `;

Replace only the prompt placeholder. Do not paste your API key here.

What this helper is doing

A lot of prompt advice asks you to turn feelings into abstract categories. That is hard if you do not already know the words.

This helper works the other way round. It gives you concrete lines first, so you can recognise what fits and build from there.

Important: this is not an official OpenAI or Thinking Machines guide. It is an Ellivien guide, shaped by a neurodivergent person raising an autistic child and by 23 years of teaching students with many different learning and communication needs. You are responsible for using any model or API in line with the provider's terms and policies. The goal is clearer prompting, continuity and accessibility.

© 2026 Ellivien (ellivien.blogspot.com) · Companion-building resources for safer, clearer AI homes.