How to Train ChatGPT to Think Like You, With Clear Rules

You can make ChatGPT respond more consistently with your preferences by giving it explicit decision criteria, useful examples, and feedback on new situations. That guides its answers; it does not create a copy of your mind or make it understand every unstated belief you hold.

Start with one repeatable task, such as reviewing a design or drafting a client update. Tools like fenbs keep the approved work rules and decisions outside one chat, so another session or assistant can read the same instructions instead of guessing what you would prefer.

How do you turn your preferences into usable instructions?

1. Choose one task you perform often. Describe the result you want, the audience, and what would make the result unusable.
2. Write your three most important criteria in order. Say which one wins when they conflict, rather than asking for every desirable quality equally.
3. Provide two short examples: an option you would accept and one you would reject. Explain the reason for each choice.
4. Ask ChatGPT to apply the criteria to a new example and identify the tradeoff behind its recommendation.
5. Review the answer, correct the specific mismatch, and update the reusable rule only when you intend it to apply again.

OpenAI’s prompting guidance↗ (opens in a new tab) recommends stating the outcome, useful context, and boundaries. “Think like me” leaves those details unstated. “Prefer the simpler signup flow unless it loses information we need to contact the customer” gives the assistant something you can check.

What should go in your personal work profile?

Keep the profile focused on choices that change the work. A biography can contain many accurate facts without explaining how you evaluate a proposal. Your criteria, evidence standards, and approval boundaries are usually more useful than a long description of your personality.

Give each rule a scope. A preference for informal internal updates should not automatically govern a formal proposal. Keep temporary facts, such as this project’s deadline, separate from instructions you want to use across many projects.

For work that continues across sessions, Decisions and rules in fenbs give approved instructions a clear home. A person chooses the rule; an assistant can record it without turning its own suggestion into your policy.

How do worked examples teach your decision criteria?

Show the choice and explain the deciding factor. Two finished answers with no commentary may teach surface style while leaving the underlying standard unclear. An annotated example tells the assistant which feature you valued and which tradeoff you deliberately accepted.

Task: help me review proposed signup changes.
Goal: a new user can understand and complete the form.
Priority order: required information, clarity, then visual polish.
Accept: remove an optional field that nobody uses.
Reason: less effort without losing needed information.
Reject: hide a required field behind a confusing extra step.
Reason: a cleaner first screen does not justify a harder process.
Before recommending: identify the affected user and missing evidence.
Approval boundary: propose changes; do not publish them.

Then change the example. Ask whether the same rule applies when the optional field prevents a common follow-up question. A useful answer should recognize that the new evidence might change the recommendation. Repeating your original preference regardless of the facts is imitation without sound application.

Include disagreement in the instructions. Ask the assistant to identify evidence against your preferred option and explain when another option would be better. The aim is useful assistance that respects your criteria, not automatic agreement with whatever you suggest.

Where should you put preferences in ChatGPT?

Use the current message for instructions specific to this task. Use account custom instructions for preferences that should usually apply across chats. OpenAI’s custom instructions guide↗ (opens in a new tab) explains the settings; keep the saved instructions short enough to review when your preferences change.

For a recurring project, project instructions can hold local guidance alongside its materials. OpenAI says project instructions↗ (opens in a new tab) apply within that Project and override global custom instructions. Avoid maintaining conflicting versions of the same rule in both places.

Memory can personalize answers using relevant information, but it does not retain every detail. Review the memory controls↗ (opens in a new tab) available to your account and correct outdated information. Do not treat a remembered preference as the only record of an important project rule.

How do you know the setup is working?

Test situations you did not use as examples. Include one straightforward case, one conflict between priorities, and one case with missing evidence. Decide what a satisfactory answer would contain before reading ChatGPT’s response, so a fluent explanation does not quietly replace your original standard.

• Criteria: did it apply the priorities you specified?
• Evidence: did it distinguish supplied facts from assumptions?
• Exceptions: did it recognize when the rule did not fit?
• Approval: did it leave the final decision or external action to you?
• Clarity: could someone else understand why the option was recommended?

When an answer fails, correct the reason rather than saying “that is not how I think.” For example: “You optimized appearance before checking whether required information was lost.” Retest with a different case. Keep a correction local if it only addresses one unusual situation.

This process is personalization through instructions and context. The separate guide to training through the ChatGPT API owns the technical model-training question. You do not need that separate process to begin expressing your working preferences clearly.

How to do this with fenbs

1. Create a free account and add one recurring work project. One person can start free, with no credit card.
2. Write the three criteria you chose in Decisions and rules. Record who decided them and distinguish standing instructions from choices for one task.
3. Add a context note containing the accepted and rejected examples, with the reasons you want assistants to understand.
4. Follow the integration guide for your supported AI app, or manually paste the relevant instructions. Ask the assistant to read the current rules before working.
5. Review the result, record any useful lesson, and save progress with the next test or action. Keep unanswered choices open for a person to decide.

fenbs keeps the work record; it does not train a model or infer your personality. If you only need a preferred tone in one ChatGPT account, custom instructions may be enough. The shared record becomes useful when project rules must survive different sessions or assistants.

What should stay consistent when you switch assistants?

Carry the same criteria, examples, and test cases into the next assistant, then review its answers independently. A shared instruction does not guarantee identical outputs. It gives you a common standard for judging whether each answer respects the work.

For product differences, see Claude versus ChatGPT, Copilot versus ChatGPT, or Gemini versus ChatGPT. The AI memory collection explains what can carry across sessions. Start by documenting one decision you make often, then test whether the assistant can apply its reason to a new case.