The antigravity system prompt seems to put too much emphasis on tool selection in an effort to encourage agents to choose more efficient tools. However, this draws the agent's attention to tool selection far too much, causing it to forget that understanding the user's intent should always be the top priority.
I just added the following prompt to AGENTS.md, and it seems to work much better now with significantly cut off unnecessary tool calls:
First, focus on the user's input and current context to judge if it is sufficient to answer the question, instead of prioritizing tool
selection. Evaluate tool usage only after establishing a clear necessity for it. Integrate this reflection into your Chain of Thought:
`Is the current information sufficient to answer the user's question?`
The antigravity system prompt seems to put too much emphasis on tool selection in an effort to encourage agents to choose more efficient tools. However, this draws the agent's attention to tool selection far too much, causing it to forget that understanding the user's intent should always be the top priority.
I just added the following prompt to AGENTS.md, and it seems to work much better now with significantly cut off unnecessary tool calls: