mirror of
https://github.com/NikkeTryHard/zerogravity.git
synced 2026-04-25 07:05:58 +03:00
[GH-ISSUE #35] Pass system prompt as rules #31
Labels
No labels
bug
enhancement
enhancement
notice
pull-request
wontfix
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
starred/zerogravity#31
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @LibertX on GitHub (Feb 20, 2026).
Original GitHub issue: https://github.com/NikkeTryHard/zerogravity/issues/35
What do you want?
Hello !
Per https://github.com/NikkeTryHard/zerogravity/issues/28, you said it's not possible to remove system prompt for stealth reasons.
An alternative I use on the official Antigravity client is to put the system prompt in
~/.gemini/GEMINI.md(global rules files) and it works pretty much as if it was a real system prompt.Could you implement it so system prompt passed by clients (eg. Open WebUI) are implemented as rules ?
Thanks
Use Case
Making system prompts works better (eg. jailbreaks).
Alternatives Considered
No response
@NikkeTryHard commented on GitHub (Feb 20, 2026):
nice idea i will reverse engineer further and see what i find. thx for the suggestion!
@DarKWinGTM commented on GitHub (Feb 20, 2026):
I can confirm that on Claude Code + CCS (Claude Code Switch), the system prompt / rules / context / skill mechanism works seamlessly without any context loss.
Here's how the environment is structured:
~/.claude/rules/*.md): Loaded as global instructions — currently running 21 rules files covering everything from anti-hallucination, document control, security boundaries, to output safety/commit,/sync-memora,/ui-ux-pro-max) are loaded and executable on demandThe key architectural insight is that rules files act as a functional equivalent to system prompts — they define the AI's behavioral constraints, persona, and operational standards. On Claude Code, this is achieved through
~/.claude/rules/(global) and.claude/rules/(project-specific) directories, which are injected as part of the context window.This approach effectively separates the environment definition (rules/context/prompt) from the conversation flow, allowing full customization without losing any behavioral context.
If zerogravity could implement a similar pattern — passing client-provided system prompts as rules/context rather than trying to inject them into the system prompt field directly — it would likely achieve the same result: full behavioral control without the stealth/detection tradeoffs.
@MIKUSCAT commented on GitHub (Feb 20, 2026):
good idea！i will also notice it
@m1a8h7sg90i commented on GitHub (Feb 21, 2026):
+1 Love this idea, the antigravity system prompt is huge, it lobotomizes Opus from my experience so this would be excellent improvement
@NikkeTryHard commented on GitHub (Feb 22, 2026):
Implemented in v1.3.3. System prompts are now injected as
<user_rules>-- the same format the LS uses for GEMINI.md rules. Models treat this as legitimate user configuration and obey it, unlike the old<SYSTEM_INSTRUCTION_OVERRIDE>tag which they detected as prompt injection.Duplicate of #42 -- both fixed by the same commit.
@MIKUSCAT commented on GitHub (Feb 22, 2026):
respect and appreciate
@terryops commented on GitHub (Feb 23, 2026):
@NikkeTryHard Now it works, but not in a perfect way. The thing chain is very much displayed and sent to the user while using openclaw, and also it not only focus on the last question, but also the last several sentences
@NikkeTryHard commented on GitHub (Feb 23, 2026):
open a separate issue as i need to reproduce it. follow the bug report format