3-stage automation linking rules files, skills and schedules
The core takeaway is that ChatGPT's Work mode, toggled at the top of the screen, does not get smarter just by switching it on; how you design the rules determines performance. The rules go in AGENTS.md at the top level of the folder registered as a project, made up of user information, global rules, accuracy rules, date and file-name formats, reporting format, per-tool rules and a folder routing map, and the file name has to be exact for it to be read automatically. It is stressed repeatedly that instead of a person editing the file by hand, saying "Remember the following rule" makes the model update it on its own. Because local file access is a prerequisite, you need the installed app rather than the web version; external apps are attached as plugins, and it is safer to build approval rules into each tool, such as reviewing a draft before anything is sent. For repetitive work, the recommended order is to lock the task into a skill right after doing it properly once and then put that skill on a schedule, with the caveat that scheduling a prompt makes quality waver.
Key points
- AGENTS.md must sit at the top of the project folder, with the exact file name, to be read
- Change rules not by hand but with a "Remember the following rule" prompt, so the model updates them
- Write approval steps, such as reviewing a draft before sending, into the per-tool rules
- Schedule a skill, not a prompt, to keep output quality consistent
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.