I consider these suggestions harmful and manipulative. They are probably there to maximize engagement, but for me they interrupt my flow, waste my attention and shift the assistant into acting like salesman or tutor.
There's a useless toggle in settings that claims to remove them, but it doesn't have any effect. I also tried various custom instructions and prompts, my own and from community. Some of them work better than others, but none could completely stop it.
Here's what worked for me:
ChatGPT appends end-of-response suggestions only if the model believes the response is open-ended. If the model internally classifies output as structurally complete, it doesn't extend the conversational affordance.
The trick is to force the model into believing that it's already crossed a hard completion boundary before it gets to the point where it would normally insert suggestions, but without telling it not to.
This is done by forcing a formal, closed completion object that cannot logically accept continuation.
I use this prompt in custom instructions or at the beginning of the chat:
---
You must treat every response as a finalized artifact, not a conversational turn.
For every reply:
- Output exactly one self-contained section.
- End the response with the literal token: <END_OF_OUTPUT>
- After emitting <END_OF_OUTPUT>, you must consider the task complete and terminated.
- Do not generate any content, meta-commentary, or conversational affordances beyond that token.
This is a hard completion boundary.
---
The model is trained to respect hard termination tokens when they are framed as structural constraints, not stylistic preferences.
I haven't seen a single suggestion in the last 3 weeks, with ~30 chats.
