Enhanced Conversational Control with In-Thread Prompt Editing & History

[left briar]

Currently, users can only edit their most recent prompt in Gemini. This proposal suggests enhancing user control and conversational flexibility by introducing two key features:

Edit Older User Prompts:

Functionality: Allow users to go back to any of their previous messages (prompts) within an ongoing conversation and modify them. Upon editing an older prompt, Gemini would then regenerate its response and subsequent parts of the conversation based on this corrected/updated user input.

Benefit: This would be invaluable for correcting typos, clarifying instructions, or refining the direction of a conversation without having to abandon the entire thread and start over. It's particularly useful in long, iterative discussions where an early imprecision can lead the AI astray.

Prompt Edit History ("Cycle Through Edits"):

Functionality: When a user edits one of their prompts (either the most recent or an older one), Gemini would save a version history of that specific user message. Users could then "cycle through" or view these past versions of their own edited prompt.

Benefit: This allows users to track their own thought process, revert to a previous version of their prompt if an edit doesn't yield the desired outcome, and easily compare the impact of different phrasings or instructions they've tried for a particular conversational turn.

[quick meteor]

Like branching storylines, I second this. Model drift is an ongoing issue and that would give the user more control as well as explore different chains of thoughts within the same conversation while not cluttering the log with irrelevant side-quest data.
Taking that concept of user control a step further, I propose full CRUD (Create, Read, Update, Delete) capabilities on all parts of the conversation - including replies and 'thought' processes by the assistant. This would not only improve readability but provide explicit, direct and high-quality feedback to the active session's context. Instead of saying 'You were wrong', the user shows the model 'This is what the answer should have looked like.', shifting from confrontational correction to constructive guidance.

[left briar]

Yeah, LM Studio has this feature and I'm surprised that the Gemini app doesn't yet have it.

[quick meteor]

🧛
I've been playing around with Silly Tavern as frontend UI (and NovelAI or OpenRouter as model provider) for a while now, not being able to edit a lot of user prompts (and model output) within the Gemini app feels weird.