I built a role-based LLM workflow for coordinating humans, LLMs, and coding agents
I built a role-based LLM workflow framework for coordinating humans, LLMs, and coding agents without losing human judgment.
As AI takes on more work, I felt that what humans need to judge must become even clearer.
So instead of treating LLMs and coding agents simply as “code generation tools,” I tried to create a workflow where:
- the human sets the direction
- the LLM organizes the scope and instructions
- the coding agent executes or investigates
- the LLM interprets and records the result
- the human reviews and makes the final judgment
I’m not sharing this as a groundbreaking invention.
I’m sure many developers have thought about similar problems, and some may already be using similar workflows.
But I built this from my own experience as a student developer who values engineering fundamentals, development logs, and keeping context throughout the development process.
I have always cared about development logs and keeping track of why decisions were made, so I wanted the workflow to preserve context instead of just speeding up implementation.
The goal is not to avoid AI.
The goal is to keep human engineering insight and technical understanding at the center, while making AI-assisted development more structured, reviewable, and practical.
I have applied this flow to my own work and organized it into a first usable version. I felt it was practical enough to share, so I made it public.
I’d like to know whether this kind of structure feels useful, too rigid, or similar to something you already use.
GitHub:
https://github.com/bfdcoco/dev-workflow-agent-en
Note: The GitHub documentation includes a GPT link configured to perform the LLM role in this workflow. To use that GPT properly, you need to be signed in to ChatGPT. If you are not signed in, some commands may appear to respond, but the intended output format and workflow can break.
The documents and public materials are released under the CC BY 4.0 license. If you use or reference them, please credit the original author and include the GitHub repository link.