Every few months a screenshot goes around with "secret ChatGPT codes" that supposedly unlock a hidden mode, a better model, or a private developer menu. I have opened a lot of those posts. I have never found an official list of secret codes, because OpenAI does not publish one.
What does exist is more useful and less exciting: ordinary product features, documented in the ChatGPT Help Center. They do not look like cheat codes. They look like settings, folders and toggles. If you want better answers, start there.
The myth, quickly
The usual claim is that a special phrase — a role, a token, a "DAN" prompt, a string that looks like a command — will bypass limits or turn the assistant into a different product. Sometimes the post is just marketing. Sometimes it is a jailbreak attempt, which sits in a grey area I will not help with.
Two practical problems show up even if you only care about getting work done:
- The phrase often does nothing, or it does something once and then stops working after a model or policy update.
- It trains you to hunt for a hidden switch instead of stating the task clearly.
If a capability is real, it usually has a name in the interface and a help article. That is the test I use.
Custom instructions: say the defaults once
Custom instructions let you tell ChatGPT, once, how you want it to talk to you: your work, your preferences, and what to avoid. After that, you do not have to repeat "be concise" and "I write firmware in C and Python" in every new chat.
I keep mine short:
- who I am and what I build
- the languages and tools I actually use
- a request to flag uncertainty instead of inventing a citation or an API
I also say what not to do: no fake benchmarks, no made-up client stories, no padding. If your instructions become a page long, the model has to weigh them against the current question, and the result gets messier. A tight paragraph is enough.
These are account-level defaults, not a secret personality. You can change or clear them at any time.
Projects: keep the context in one place
Projects in ChatGPT are the feature I wish more people used. A project holds related chats, files and instructions together, so you can come back to a piece of work without re-explaining the background.
That sounds small. It is not, if you are writing firmware for one device, drafting a series of articles, or debugging the same service for a week. I put the notes, a couple of source files, and a short project instruction such as "this is the EnergyLock firmware; prefer the existing naming; do not invent hardware that is not in the files."
Two memory modes matter, and OpenAI documents both:
- Default memory can still use your wider ChatGPT memory, if memory is on for your account.
- Project-only memory keeps the project more self-contained. Chats in the project can refer to other chats in the same project, but not to your life outside it.
You can choose the setting when you create a project and, as of mid-2026, change it later in the project settings. If you share a project with other people on a plan that supports sharing, OpenAI says project-only memory turns on automatically and stays on. That is a feature, not a bug: it reduces the chance that a personal note leaks into a shared workspace.
Do not drop client secrets, passwords or private keys into a project just because it feels like a folder on your disk. It is still a conversation with a hosted service.
Memory: useful, and worth inspecting
ChatGPT can remember facts you want it to keep, such as "I prefer Python examples" or "I am based in Lithuania." OpenAI documents this as a setting you can manage, not as a hidden profile.
I treat memory as a list I review, not as a diary. I look through saved memories now and then and delete anything that is stale or too personal. If I am about to discuss something I do not want remembered, I use a temporary chat instead of hoping the model will "just know" to forget it.
Temporary chat: a clean room for one task
Temporary chat is the official way to have a conversation that stays out of your usual history. OpenAI's help page is the source to trust for the current rules, because details (including how personalization and memory interact) have changed as the product has evolved.
The idea is stable enough to use: if the topic is sensitive, experimental, or just something you do not want mixed into later chats, start a temporary chat. It is a documented button, not a prompt you have to memorise.
Temporary chats are also a good place to test a prompt before you drop it into a long-running project.
Data controls: one screen, several decisions
Data controls are easy to skip because they sit in a settings menu rather than in the chat box. They are not a "secret," but a lot of people never open them.
Depending on your plan and region, this is where you decide whether eligible conversations can be used to improve models, how chat history works, and how to export or delete your data. I will go through the practical checks in a later article, because the details differ between a personal account and a work plan. The short version: open the menu once, read the official wording, and do not assume the default matches what you want.
Tools you already have, if your plan includes them
Depending on the plan, ChatGPT may also give you voice, image generation, file uploads, web search, or a canvas-style editor for longer drafts. Availability changes, so I will not list every button as if it were permanent. The useful habit is the same: if you need a tool, look for it in the product and in the help center, not in a viral list of codes.
What I use instead of "secret codes"
When I want a better answer, I do four ordinary things:
- Write a clear prompt: goal, format, constraints, context.
- Put repeating context into custom instructions or a project, not into folklore.
- Use a temporary chat when I do not want the thread mixed with everything else.
- Check data and memory settings so the product is not quietly keeping more than I intended.
None of that is hidden. It is also more reliable than a phrase copied from a thread that will be outdated next month.
I'm Saulius, an independent full-stack, automation and electronics developer in Kaunas, Lithuania. You can also find me on Dev.to and LinkedIn.
Top comments (0)