I’m building a ChatGPT-based agent that is supposed to follow a recurring workflow. I’m using ChatGPT Pro.
ChatGPT itself suggested that we use a text-based control document, stored locally, as the “source of truth.” It defines the workflow, rules, and important decisions. ChatGPT also updates this document as we develop or modify the system together. The ideia is that i can control and tweak the process, at least that was the goal.
In simplified form, the workflow consists of:
-
Import information from multiple sources
-
Read, interpret, and structure the information
-
Perform research and analysis
-
Create a structured working document
-
Process and quality-check the information together with the user
-
Export it into a final document and verify the result
The control document also divides the process into clearly defined steps, with rules specifying what must be completed before moving on to the next step.
The problem: For example, I can say, “Start a new project and complete steps 1–4 according to the control document.” Even though ChatGPT has access to the document and has itself helped write and update it, it will sometimes complete only parts of the steps, skip a rule, or stop the process too early.
I’m therefore trying to understand whether the architecture itself is wrong. Is a text-based control document the wrong way to manage a longer agent workflow? And if so, what are better alternatives? (it doesn´t have to be chatGPT)