AIA proposed agents.md rule for separating casual chat from task instructions
A user recommends adding instructions to agents.md intended to keep GPT 6 models from pausing work over casual messages. The proposed rule distinguishes conversation from messages that start, change, cancel or continue work.
TLDR
The user’s suggested wording treats only messages that start, change, cancel or continue work as “steering.” Other messages answerable from current context should get a normal answer—or a single reaction when sufficient and unambiguous—without triggering tools or task-state changes solely because of the message. In a September 25, 2026 follow-up, the user says the wording is apparently from a leaked system message for an upcoming product.
Combined views
6.3K
2 Sources, first seen 18h ago
A proposed agents.md rule for separating casual chat from task instructions
A user recommends adding instructions to agents.md intended to keep GPT 6 models from pausing work over casual messages. The proposed rule distinguishes conversation from messages that start, change, cancel or continue work.
TLDR
The user’s suggested wording treats only messages that start, change, cancel or continue work as “steering.” Other messages answerable from current context should get a normal answer—or a single reaction when sufficient and unambiguous—without triggering tools or task-state changes solely because of the message. In a September 25, 2026 follow-up, the user says the wording is apparently from a leaked system message for an upcoming product.