Four ways AI drifts
Drift is not one thing. It helps to know which one you are fighting.
- From the brand: the text and the pictures look like somebody else's.
- From the topic: the answer slides away from what you actually asked.
- From the thread: it forgets what you settled three messages ago.
- From the project: campaign A leaks into product B.
Why it happens
A language model only sees what is in the conversation right now. In a long thread, the first messages get diluted by everything that came after, and in the end they drop out. A new chat starts from zero. And if you use one chat for three projects, the model has no way of knowing which one you mean.
Custom instructions help a little, but they apply everywhere. They cannot say that this campaign has one tone and that product another.
What helps in any tool
- One conversation per topic. Start a new one when you switch tasks.
- Keep projects apart. Use the tool's projects or workspaces, not one long chat.
- Put a short brief at the top: who, for whom, which tone, what must never be said.
- When a thread gets long, ask for a summary of what you have decided, and start again from it.
Why the drift keeps coming back
Every fix above has one thing in common: you are the memory. You have to remember to paste the brief, update it when something changes and keep the projects apart yourself. The day you forget, it drifts again.
What a drift-proof setup looks like
The brand sits at the root, with its voice, audience, visual direction and rules. Projects sit underneath and inherit the brand, but keep their own facts, so they never mix with each other. Instructions you give are saved on the project they belong to. And every answer starts from the right brand and the right project, without you pasting anything.
In Marshi
In Marshi, work lives in your brand and its projects. Child projects inherit the voice, audience and rules, saved instructions stay with the project, and a chat that turns into real work moves into a project with its whole history.
