Issue
The “super” system and user prompt applied to every agent session (including sessions that use custom agents) is very large. Because it is always injected, it can bloat the context even when major sections are not relevant to the current task.
Users should have better visibility (besides the chat debug view) and control over this “always-on” prompt so they can remove unnecessary context and potentially improve the agent’s output quality and consistency.
Proposed solutions
-
Direct access to the prompt source (most powerful, but potentially overcomplicated)
- Expose the full global system/user prompt in a prompt file (or equivalent configuration) that users can open, review, and modify.
- Provide a clear mechanism to manage variations across custom agents, sessions, and workspaces (including an easy way to revert to defaults).
-
Configurable prompt sections via custom agent headers (simpler and more intuitive to use)
- Add properties in custom agent headers that can enable/disable predefined chunks of the global prompt.
- Chunks could map to individual XML sections of the system/user prompt, higher-level groupings, or a hybrid approach similar to how tools are organized into groups. (recommendation: hybrid)
The goal is to reduce prompt bloat, improve effective context utilization, and give users predictable control over the agent's context in each session.
Issue
The “super” system and user prompt applied to every agent session (including sessions that use custom agents) is very large. Because it is always injected, it can bloat the context even when major sections are not relevant to the current task.
Users should have better visibility (besides the chat debug view) and control over this “always-on” prompt so they can remove unnecessary context and potentially improve the agent’s output quality and consistency.
Proposed solutions
Direct access to the prompt source (most powerful, but potentially overcomplicated)
Configurable prompt sections via custom agent headers (simpler and more intuitive to use)
The goal is to reduce prompt bloat, improve effective context utilization, and give users predictable control over the agent's context in each session.