AI viboors Those of you who ask how I use AI. This repo has every prompt I used to build the repo committed and easy to read and unlike other repos is mostly AI generated 1. I prepare a base prompt in a .claude command that has context I want always included at start of prompt 2. Base prompt + my messy brain dump of requirements generate the plans in prompts/ folder 3. I skim the plan to make sure it makes sense ask for changes potentially 4. I deploy fresh agents with fresh context windows over and over again to make progress on the plan 5. I validate tests pass and API looks good at end

Base prompt. The base prompt serves two purposes 1. Getting context into MY head 2. Getting context into the LLM

Generate a large plan. I commit the plan so it's easy to update and regenerate code Important: I generate tests and APIs before implementing anything. This creates yet another easy to review point if I choose to review

Before I would only run plans serially but now I utilize subagents with a dispatcher agent whose job it is to figure out what tasks can be done in parallel which greatly speeds up how fast I get back code results

The goal of the workflow is pretty simple 1. Make it fast and easy to review what the AI is doing 2. Force the AI to gather context seperately from executing so the executing context window is clean 3. Make it easy to flag a plan, api, or execution as not good enough, update the plan, and retry again These three concepts we can condense to 1. Optimize review bottleneck 2. Manage context 3. Fail fast and iterate