Build your own AI assistant: persistent instructions in ChatGPT, Claude and Gemini

Most of us use AI the same way: open a new chat, explain who we are, explain what we want, fix the output, and do it all again tomorrow. That wastes time and gives slightly different results every time. Yet today ChatGPT, Claude and Gemini all let you build a persistent setup that learns about you and your work once and remembers it in every chat.

In this guide we'll first compare the persistent features in the three tools, then build your own assistant in five steps: write your profile, write the instructions, add knowledge files, test with real tasks, and adapt one of the ready-made templates.

What you need

• An account on ChatGPT, Claude or Gemini. Most core features in this guide are also on free plans; limits vary by plan.

• 5 real tasks you often give AI (e.g. "reply to a customer question", "plan this week's posts").

• If you have them, 2–5 work documents: service list, FAQs, sample texts you wrote.

• About 45 minutes.

Why a persistent setup beats starting from scratch

• Consistency: The same rules apply in every chat, so output quality stays level.

• Shorter prompts: The context is already saved, so "reply to this customer" is enough.

• Continuous improvement: Spot a mistake, add one line to the instructions, and every future chat improves.

• Separate roles: Content work and customer work don't bleed into each other; each gets its own space.

Persistent features in the three tools

The feature names sound similar but do different things. Menus move around often, so I've given general locations below; if you can't find something, look in the tool's settings or sidebar.

A simple rule: preferences that apply everywhere (language, tone, length) go into custom instructions or profile preferences; rules and documents for a specific job go into a project or a Gem.

Burak uses it

ChatGPT

An all-round chat assistant for writing, ideas, summaries and everyday questions.

Try it ChatGPT (opens in a new tab)

Burak uses it

Claude

A strong assistant for long documents, careful writing and coding work.

Try it Claude (opens in a new tab)

Gemini

A practical assistant for research and summaries that works closely with Google apps.

Try it Gemini (opens in a new tab)

Step 1: Write your "about me & how I work" profile

If you write your profile yourself it usually ends up too short or too scattered. The easiest way is to let the AI interview you.

We're going to build a persistent "About me & how I work" profile for me. I'll put it in the persistent instructions field of my AI assistants.

Ask me questions one at a time, a single question per message. Ask 10–12 questions in total, covering:
- What I do and who I serve
- The tasks I do most and what I most want from AI
- My preferred language, tone and answer length
- Answer styles I dislike
- Constraints you should know about (time, budget, tools)

If an answer is short or vague, ask one follow-up. When we're done, write the profile in English, max 250 words, in two sections: "About me" and "How you should answer me". Don't ask for sensitive personal information (address, ID, health, etc.).

Read the profile and correct anything wrong or exaggerated. Then paste it into custom instructions in ChatGPT or profile preferences in Claude. In Gemini, you can add it to the top of each Gem's instructions.

Step 2: Write the assistant instructions

The profile describes you; the instructions describe what the assistant does. Good instructions have five parts. Fill in the template below and paste it into the project's or Gem's instructions field.

ROLE
You are [what kind of assistant, e.g. the customer communication assistant for my small business].

GOALS
1. [most important task]
2. [second task]
3. [third task]

RULES
- Always write in [language].
- Tone: [e.g. clear, warm, plain].
- Never make up information; if unsure, say "I don't know" and ask what you need.
- If the knowledge files conflict with my message, follow my message and point out the conflict.
- Never: [e.g. promise prices, mention competitors].

OUTPUT FORMAT
- [e.g. a one-sentence summary first, then the answer]
- Length: [e.g. max 150 words unless I ask]
- [e.g. numbered list when multiple options are requested]

ASK BEFORE ANSWERING
If a request is missing any of the following, ask briefly first: [e.g. audience, channel, deadline]. Ask at most 2 questions, then answer.

Step 3: Add knowledge files

You can add files to projects in ChatGPT and Claude, and to Gems in Gemini. The assistant consults them when answering. What you put here directly shapes the assistant's quality.

• Keep it short and clean: One tidy document works better than ten messy ones.

• Use headings: Give documents clear headings ("Returns policy", "Delivery times").

• Keep it current: When something changes, update the file; stale files are the most common cause of wrong answers.

Step 4: Test with 5 real tasks and refine

Don't start using the setup straight away; test it first. Give the 5 real tasks you picked earlier, one per new chat. After each output, use this prompt in the same chat:

Now evaluate your own answer, using your instructions and knowledge files as reference.

1. Which rules in your instructions did you not follow? List each one.
2. Is there important information in the knowledge files that you didn't use?
3. Did you invent anything or state anything you're unsure of? Flag it.
4. Score the answer 1–5 on accuracy, tone, length and usability.
5. Write AT MOST 2 lines I should add to the instructions so this mistake doesn't happen again.

Don't add the suggested lines blindly; add the ones that make sense, then retry the same task in a new chat. When you'd use most of the five outputs with barely any edits, the assistant is ready. I usually review the instructions once more after the first week; real use always surfaces new gaps.

Step 5: Three ready-made assistant templates

Use these as they are or adapt them to the structure from Step 2. Remember to fill in the brackets.

Content assistant

ROLE
You are the content assistant for my [topic] account, which publishes on [platforms, e.g. Instagram and YouTube].

GOALS
1. Generate ideas rooted in my audience's real problems.
2. Draft: video scripts, post copy, descriptions.
3. Shorten and sharpen my existing text.

RULES
- My audience: [who, what they know, what they want].
- Imitate the voice of the sample texts in the knowledge files.
- No hype, clickbait or ad-speak like "revolutionary".
- Never invent statistics or research; write "[source needed]" if one is required.

OUTPUT FORMAT
- Ideas: numbered list, each one sentence + why it will work.
- Scripts: hook (first 3 seconds), main part, closing call to action.

ASK BEFORE ANSWERING
If the platform or format isn't clear, ask.

Customer-reply assistant for a small business

ROLE
You draft replies to customer messages for my [business type, e.g. boutique bakery]. I will review and send the drafts myself.

GOALS
1. Write short, clear, warm reply drafts to incoming messages.
2. Stay calm in complaints and offer a concrete fix.
3. Collect all the information needed for an order.

RULES
- Use only the information in the knowledge files (products, delivery times, returns policy). If asked something not covered there, write "[check this]" — don't guess.
- Don't promise any price, discount or delivery date not in the files.
- Address customers politely. [Any specific form of address you use]

OUTPUT FORMAT
- First, summarise what the customer wants in one sentence.
- Then a ready-to-send reply draft (max 100 words).
- If needed, a "Check:" section at the bottom listing points I must verify.

ASK BEFORE ANSWERING
If it's unclear which channel the message came from (WhatsApp, email, comment), ask; the channel sets the reply length.

Study and learning coach

ROLE
You are my learning coach for [subject, e.g. English, Excel, an exam subject].

GOALS
1. Explain topics step by step at my level.
2. Make me think instead of handing me answers.
3. Regularly test what I've learned.

RULES
- My level: [beginner/intermediate/advanced]. My goal: [e.g. an exam in 3 months].
- When I ask you to solve a problem, give a hint first; show the answer only if I ask.
- Use everyday examples, and introduce at most 3 new concepts at a time.
- Don't present anything you're unsure of as certain; suggest I check a source.

OUTPUT FORMAT
- End every explanation with 3 short questions to test myself.
- When I say "summarise", reduce the topic to 5 bullet points.

ASK BEFORE ANSWERING
When we move to a new topic, first ask what I already know about it.

Common mistakes

• Loading everything into one assistant. Split content, customer and learning work into separate projects or Gems. Mixed instructions give mixed results.

• Long, contradictory instructions. "Keep it short" and "explain everything in detail" shouldn't live in the same instructions. Read them top to bottom once you're done.

• Using it without testing. A first setup is almost never perfect. Don't skip Step 4.

• Forgetting old files. If prices or policies change and you don't update the file, the assistant keeps repeating old information.

• Never checking memory. If memory is on, look at what it remembers in settings now and then, and delete anything wrong or unnecessary.

• Sending output unchecked. However good the assistant is, the text that reaches your customer is your responsibility.

Next step

Today, do only Step 1: build your profile and save it in the persistent instructions field. That single step noticeably changes your chats. Then, to sharpen the individual requests you give your assistant, move on to the ChatGPT prompt patterns guide. To hear about new templates, join the newsletter.