My raw workflow for brainstorming a new Learning Mission.
The Plan.md method
I kept losing the plot when trying to build Google Gems and LLM projects. The AI would hallucinate a structure or forget my constraints halfway through, and I needed a single base document to keep the whole workflow from derailing.
The Stack
A blank plan.md file, and a handful of other markdown files routed within the plan.
The Inputs
A plain-text outline of the exact workflow I want the AI to execute, broken into hard, chronological parts. Plus, the specific markdown files containing my rules and constraints.
The Mechanics
I stopped trying to write massive, complicated system prompts. Now, I just create plan.md and write out the instructions like a dumb checklist.
I tell the model exactly what to do: Part 1, break the idea down. Part 2, follow my specific workflow template. Part 3, edit the text strictly using my rules to strip out generic AI slop. I feed that plan file into the system first. It forces the model to read the map and understand the routing before generating a single token.
The Output
A rigidly structured AI session that actually outputs a final, usable asset. The plan.md stays pinned as the base document, acting as the brain for the entire project so I never have to repeat myself.
Phil