ChatGPT custom instructions drift from your brand because they are a block of text someone typed once, and a brand changes faster than anyone remembers to update that text. A marketer pastes the voice guidelines into the custom instructions field in January. By spring, a word gets added to the banned list and a color value shifts, and ChatGPT never hears about either change. This post names why that drift happens and what keeps ChatGPT's output on-brand instead: a design brief that stays current and reaches the AI directly, through an MCP connection.
Key Takeaways
- ChatGPT custom instructions are a static field in Settings, read on every chat but never updated except by a person going back to edit it.
- Most teams write custom instructions once and never revisit them, so the brand in the field and the brand in the current guide pull apart within a few months.
- TrueKeep keeps a design brief current and serves it through an MCP connection, so Claude, ChatGPT, and Cursor read the brand's rules before they write, not from a pasted paragraph.
- A final pass checks AI output against the brand and fixes what is off before it ships.
- The brand kit skill carries the same brief into Claude as a downloadable skill, for teams who want it inside a Claude conversation specifically.
What ChatGPT custom instructions actually do
Custom instructions are a text field in ChatGPT's Settings, under Personalization. Whatever is typed there gets added to the model's context on every new chat, on the web, desktop, and mobile apps. OpenAI's own help article on the feature is clear that it only changes future conversations: edits never reach a chat already in progress, and nothing updates on its own (OpenAI, "ChatGPT Custom Instructions").
That is the whole mechanism. It is a note, not a connection. ChatGPT re-reads the note every time, but the note itself only changes when a person opens Settings and retypes it.
Why custom instructions drift from your brand
The drift has a simple cause: the field is copied from a brand guide at one point in time, and the guide keeps moving. A style guide gets a new banned word. A color value changes after a redesign. A rule about how formal to be with customers gets rewritten after a bad email goes out. None of that reaches the custom instructions field unless someone remembers it exists and goes back to edit it.
In practice, nobody owns that edit. The brand manager who wrote the guide update does not know a colleague pasted an old version into ChatGPT eight months ago. The colleague has no reason to check back. The result is two sources of the brand, not one, and they disagree more with every update that only reaches the guide.
A design brief instead of a pasted paragraph
A design brief is the written-down version of a brand's decisions: its tone, the words it avoids, its color values, how formal it is with customers. TrueKeep keeps one design brief for the whole portal, and curators update it in edit mode whenever the brand changes. A brief excerpt might read like this:
voice:
tone: quiet, precise, confident without swagger
avoid:
- "seamless"
- "revolutionize"
- "best-in-class"
colors:
primary: "#1B3A34"
That file is what reaches Claude, ChatGPT, and Cursor through the MCP connection, not a paragraph someone remembers to retype. When the brief changes, every connected AI tool reads the new version on its next run. Nobody has to find the update and paste it anywhere, because the guidelines come to the work.
ChatGPT custom instructions vs an MCP connection
Custom instructions and an MCP connection solve different problems. Custom instructions are a personal setting: one person's preferences, stored once, attached to their account. MCP, the Model Context Protocol, is the open standard Anthropic introduced for connecting an AI tool to a live source instead of a static prompt (Anthropic, "Introducing the Model Context Protocol"). An MCP connection built on it is a live link between an AI tool and a source that can change: TrueKeep's design brief, read fresh on each call.
The practical difference shows up in a single question: who updates it, and how often? A custom instructions field updates only when a person remembers to open it. A design brief updates when the brand changes, and every AI tool connected to it reads the new version immediately, whether that is ChatGPT, Claude, or Cursor. For teams working inside Claude specifically, the brand kit skill brings the same brief in as a downloadable skill rather than a field to maintain by hand.
What to do with your custom instructions now
Custom instructions are not wrong to use. They are the right place for a person's own preferences: reply in a certain format, skip the pleasantries, write at a certain reading level. They are the wrong place for a brand's rules, because a brand's rules belong to the whole team, not to whoever last edited their own Settings.
A reasonable split looks like this: keep personal formatting preferences in custom instructions, and keep brand rules, voice, and approved assets in a source every connected tool reads directly. That second part is what a brand kept in TrueKeep does: one portal, one current brief, read by every AI tool the team already uses.
The honest tradeoff: someone still has to keep it current
A design brief fixes the update problem, but it does not remove the need for an owner. A brief that nobody edits drifts from the brand just as a custom instructions field does, only more slowly, because more people are reading from the same place. The difference is that one edit in TrueKeep reaches every connected AI tool at once, instead of needing to be copied into every teammate's Settings by hand.
The work still has to happen. It just happens once, in one place, instead of once per person per tool. That matters for more than tidiness: UX researchers at Nielsen Norman Group have found that a brand's tone of voice measurably shapes how trustworthy and friendly readers judge it to be, so a drifted voice is a real cost, not just an awkward one (Nielsen Norman Group, "The Impact of Tone of Voice on Users' Brand Perception").
FAQ
What are ChatGPT custom instructions? ChatGPT custom instructions are a text field in Settings that gets added to every new chat's context. They cover personal preferences, like tone or format, and apply only to future conversations, not ones already in progress.
Can custom instructions replace brand guidelines? Not reliably. Custom instructions are a static paragraph that only changes when someone edits it by hand, so they fall out of date as soon as the brand guide moves and nobody updates the field to match.
Why does ChatGPT forget my brand voice between conversations? If the voice lives only in one person's custom instructions, it never reaches a teammate's account, a different AI tool, or a chat started before the field was filled in. A source every connected tool reads directly, instead of a personal setting, solves that.
How do I give ChatGPT my brand guidelines automatically? Connect ChatGPT to a design brief through an MCP connection, so the AI reads the brand's current rules on each run instead of a pasted note. TrueKeep's AI connection guide covers the setup for ChatGPT, Claude, and Cursor.
Is an MCP connection the same as custom instructions? No. Custom instructions are a fixed paragraph stored once, in one person's account. An MCP connection is a live link to a source that can change, read fresh by the AI on every call.
Does this work for Claude and Cursor too, not just ChatGPT? Yes. The same design brief reaches any AI tool connected over MCP. TrueKeep's guide to Claude and brand guidelines covers the same mechanism from the Claude side.
Conclusion
ChatGPT custom instructions drift from your brand because they are a paragraph a person typed once, not a connection to the brand as it stands today. The fix is not a better paragraph. It is a design brief that stays current and reaches ChatGPT, Claude, and Cursor directly, through an MCP connection, with a final check on what each tool produces. Custom instructions still have a place for personal preferences; the brand's rules belong somewhere every connected tool can read. Get started to connect your brand's design brief to the AI tools your team already uses.