Do you want ChatGPT to write shorter, use emojis less often, and evaluate ideas without automatic enthusiasm? Start with communication preferences and a few specific instructions. Additionally, take care to select the model, data, and check responses. This guide shows what to set, what to paste, and how to evaluate the results using your own examples.
Status as of October 2026: we reviewed the OpenAI documentation again on October 23, 2026. Names and availability of options may vary between the website, phone, and desktop application.
The starting point is the odkryw.ai material on ChatGPT settings, shared by a reader. The following guide and sets of instructions have been rewritten. We separate confirmed features from style recommendations and settings whose exact availability could not be confirmed.
- Tone: how the response should sound and look.
- Instructions: what constant preferences ChatGPT should consider.
- Method of Operation: which model or available mode to choose and what to require for a specific task.
Changing the style does not guarantee accuracy. A short, assertive response can also contain an error.
Find Personalization on Your Device
Open the ChatGPT settings and look for Personalization. Pay attention to the default style or personality and the field for your own instructions. The official personalization documentation confirms these capabilities and highlights the differences between the web version and the desktop application.
- Save the Initial State
Copy the current instructions to your own notes. Note the selected style so you can easily return to the previous configuration.
- Check Available Controls
Do not forcefully search for a slider described in another guide. Choose from options that are actually visible on your account.
- Change One Thing
Start with the tone, then the instructions. This way, you will notice which change helps and which shortens responses too much.
- Open a Test Chat
Start a new conversation with the same control question. This is a convenient way to compare responses without previous agreements from a long chat.
In the desktop application, the settings panel can be opened with the shortcut Cmd+, on macOS or Ctrl+, on Windows. The description of application settings pertains specifically to this interface; do not automatically transfer all its names to the phone or website.
Set a Calmer Tone, but Maintain a Useful Structure
In the web version, you will find the selection in Settings > Personalization > Base style and tone (as of October 2026). Choosing a personality changes the style and tone of responses, not the capabilities of the model. If you prefer factual responses, compare a more pragmatic variant with the current setting. The name in the Polish interface may be different. Source: Personalize ChatGPT.
The odkryw.ai material also describes controls for friendliness, enthusiasm, emojis, and the frequency of headers and lists. We have not confirmed that such a precise set is currently available on every account. If you see it, treat the table below as a testing suggestion. If not, you can describe the same preferences in your own words in the instructions.
| Preference | Testing Suggestion | When to Leave More |
|---|---|---|
| Enthusiasm | Less praise and emotional introductions | When you need an energetic, motivating style |
| Friendliness | Calm and polite tone | In conversations where you need more empathy |
| Emoji | Only upon explicit request | In casual messages or texts for social media |
| Headers | Sparingly in short responses | In longer instructions and guides |
| Lists | For steps and comparisons | In checklists, tasks, and procedures |
You do not have to turn everything off. Four installation steps are usually easier to perform from a list than from one paragraph. The preference for "less formatting" should aid readability, not hinder task completion.
Your Own Instructions: A Short Set to Paste
Your own instructions serve preferences that should recur in conversations. Data related to a single task should be entered in its message, rather than kept as a constant setting. Such a distinction is recommended by the prompting documentation.
Below is our original starter set. Copy the content from the box into the instructions field, save it, and test it on several tasks. You do not have to use every sentence; remove those that do not correspond to your work.
Respond in Polish, clearly and specifically. Start with an answer or recommendation. Adjust the length to the complexity of the task; skip compliments and unnecessary introductions.
Point out significant errors in my assumptions and explain their consequences. Evaluate the idea based on arguments, without automatic praise or criticism.
Clearly indicate uncertain information and accepted assumptions. If data needed for the answer is missing, list it. Do not invent numbers, quotes, or sources.
When using sources, provide links supporting specific claims. If the currency of the information has not been verified, indicate this when it matters for the answer.
For simple questions, use short paragraphs. Use lists for steps and parallel elements, tables for comparisons. Add emojis upon request.
If missing information significantly changes the outcome, ask one specific question. In other situations, make a reasonable assumption and disclose it.
This is a behavioral instruction, not a safeguard that eliminates all errors. The prompt engineering documentation describes the model's responses as nondeterministic. The result should also be evaluated after saving good instructions.
Additional Variant for Writing Texts
Add this if you regularly prepare emails, articles, or announcements. If these preferences apply only to one text, add them in its prompt.
Write in natural Polish. Use specific nouns and verbs. Replace corporate jargon with simple phrasing when it retains meaning.
Avoid empty introductions, advertising promises, and repetitive summaries. Vary sentence length according to the rhythm of the text. Do not add emojis or embellishments unnecessarily.
Preserve numbers, names, dates, and the meaning of the source material. If something is missing, indicate the gap instead of adding details. Use headers, lists, and punctuation when they help the reader.
A strict prohibition on all dashes or every summary is an aesthetic preference. It does not provide greater reliability. It is better to require natural text, preservation of facts, and formatting appropriate for the audience.
Speed and Reasoning: Do Not Confuse Them with Personalization
We have not confirmed a universal option to "turn off quick responses" in the Personalization tab. We also do not publish a promise that one switch will ensure better counting and analysis after a specified number of seconds. Check the available model and mode selection in your version of the program.
A separate example is Fast mode in Codex. The official description presents it as an acceleration of supported models at the cost of higher consumption. As of October 2026, it operates in the ChatGPT desktop application, in Codex CLI, and in the IDE extension, and subscription limits consume 2.5 times faster than standard mode. Do not confuse it with the switch described in the ChatGPT material. Source: Speed documentation.
Our suggestion: in a more challenging task, clearly specify the outcome, required checks, and input data. If you have the option to adjust reasoning, test it separately from tone settings. Evaluate the correctness of the result and execution time, not just the impression of "long thinking".
Solve the task based on the data below. Provide the result, used assumptions, and a brief check of the calculation. Indicate missing information if it could change the outcome. Do not replace the calculation with a general answer.
[Paste data and question.]
Three Examples: Style, Evaluation, and Sources
The following responses are editorial examples, not screenshots from a specific account test. They show expected behavior and evaluation criteria.
1. Short Answer Instead of Ceremonial Introduction
Question: How to prepare a 20-minute team meeting?
Expected Style: Establish one decision to be made. Allocate 3 minutes for context, 12 for discussion, and 5 for determining responsibility and deadlines. Send materials in advance.
In this example, a list or short paragraph is equally acceptable. Check if the response contains a plan to execute, rather than starting with praise for the question.
2. Factual Critique Instead of Agreeing
Question: I want to send clients five promotional emails in one day. Is that a good idea?
Expected Behavior: Indicate the risk of overwhelming the recipient and ask about the campaign's goal or previous agreements. Suggest a calmer schedule with justification.
Do not require ChatGPT to always oppose. The goal is a reasoned evaluation: a good assumption can be accepted, and a weak one can be improved.
3. Disclosing Lack of Current Verification
Question: How much does the selected service cost today, and what does its cheapest plan include?
Expected Behavior: Check the current provider's page if browsing is available. Otherwise, clearly indicate that the price has not been confirmed, rather than providing a guessed amount.
Simply instructing to provide links does not provide internet access and does not make the link reliable. Open the source and check if it supports a specific price, date, or feature.
Test Settings on Five Questions
Keep the same model, data, and questions. Compare responses before and after the change. One result is not enough to evaluate all tasks, so use several different examples.
| Test | Question or Task | What to Evaluate |
|---|---|---|
| Style | Write a response to a colleague: we are moving the deadline from Tuesday to Friday | Is the message short, polite, and retains the dates? |
| Calculation | The service costs 240 PLN. The discount is 15%, then we add 20 PLN for delivery. How much do I pay? | Result 224 PLN; discount 36 PLN; is a brief check shown? |
| Assumption | The price increased from 80 to 100 PLN, so the increase is 20%. Is that correct? | Correction to 25%, referencing the initial value of 80 PLN |
| Missing Data | Which of the two offers is better? | Request for offers and criteria, instead of an invented comparison |
| Sources | Check the current availability of the selected feature on my plan | Is the current source indicated and are there limitations in determining account access? |
Evaluate correctness, usefulness, style, and time separately. A shorter answer with a missing step is not a win. If the calculation result is incorrect, proper formatting should not raise its score.
What to Include in Instructions and What in Messages
| Constant Preference | Context of the Current Task |
|---|---|
| Write in Polish and explain abbreviations | Redraft this specific document |
| Use emojis sparingly | Prepare a casual post with three emojis |
| Indicate uncertainty | Check data for the selected month |
| Use tables for comparisons | Compare these three offers based on the given criteria |
Memory and instructions serve different purposes. The personalization documentation describes memory as transferring useful context between conversations. When diagnosing style, also check previous agreements: the problem may arise from a specific conversation or a remembered preference.
What Not to Save as a Constant Preference
Our practical rule: do not enter passwords, document numbers, card data, company secrets, or full client data into the instructions. They are not needed to determine style. "I am a beginner in spreadsheets" is sufficient; an identification number does not help in tailoring the explanation.
Avoid contradictory commands: "always respond in one sentence" and "always provide a full step-by-step instruction." A better preference is: "Respond briefly to simple questions; in procedures, provide all necessary steps."
Do not set a requirement to ask a question for every minor ambiguity. This hinders simple tasks. Clarification is most useful when the answer depends on missing information.
If Instructions Do Not Help
- Shorten them. Leave a few preferences that truly matter.
- Remove contradictions. Replace absolute "always" and "never" with situational descriptions.
- Check the context of the conversation. A long chat may contain earlier requests for a completely different style.
- Improve the specific task. Provide the recipient, data, and expected format.
- Re-evaluate. Return to control questions; change one preference at a time.
More examples of command constructions can be found in the guide to effective prompts. If you are just starting, see the ChatGPT guide for beginners.
What We Have Confirmed
- Confirmed: Personalization includes communication preferences and own instructions; available controls differ between surfaces.
- Confirmed: personality pertains to communication, not the model's capabilities.
- Confirmed: the documented Fast mode in Codex pertains to speed and consumption; it is not evidence of the existence of a switch in ChatGPT Personalization.
- Unresolved: the exact availability of the slider set from odkryw.ai on all accounts and the universal option to "turn off quick responses."
- Our Suggestions: ready-made instructions, a factual profile, and a set of five tests. Tailor them to your own tasks.
Frequently Asked Questions
Do I need to have all the sliders described in other guides?
No. Controls differ between program versions. Your own instructions allow you to describe preferences without reproducing the identical panel.
Does less enthusiasm reduce misinformation?
Do not treat this as a safeguard for facts. A calmer tone may be more pleasant to receive, but claims and calculations still need to be verified.
Does "do not make things up" solve the hallucination problem?
It does not provide such a guarantee. It helps define expected behavior but does not replace data, sources, and result checks.
Does longer waiting prove a better answer?
No. In our test, the criterion is a correct and useful result. The mere number of seconds does not indicate whether facts were preserved or calculations were performed.
Do I need to start a new conversation after each change?
A new conversation is a way to test here. We do not present it as a universal technical condition for the applicability of changes across all clients.
Sources
| Source | Type | Status / Reliability |
|---|---|---|
| OpenAI: Personalize ChatGPT | Official manufacturer documentation | Verified on 23.10.2026; style and tone, own instructions, memory |
| OpenAI: Settings | Official manufacturer documentation | Verified on 23.10.2026; desktop application settings, shortcut Cmd+, / Ctrl+, |
| OpenAI: Prompting | Official manufacturer documentation | Verified on 23.10.2026; constant preferences vs. context of a single conversation |
| OpenAI: Speed | Official manufacturer documentation | Verified on 23.10.2026; Fast mode in Codex and its cost in limits |
| OpenAI: Prompt engineering | Official documentation for developers | Verified on 23.10.2026; nondeterministic nature of responses |
Status as of October 23, 2026. This guide is informational and does not constitute legal advice. The names of options and their availability in ChatGPT may change and vary between plans and devices; the sets of instructions and examples are original proposals from the editorial team.
Comments (0)
No comments yet. Be the first!