Source highlight:

Outputs are disposable; plans and prompts compound. Debugging at the source scales across every future task. It transforms agents from code printers into self-improving colleagues.

My next step would be to review my PRD template and extract it into a separate file. What’s the easiest way to continuously improve it? Perhaps having project-specific templates, since not all projects have the same requirements. However, plan templates can be forked and include a section similar to a local CLAUDE.md file, but stored in the repository.

I’m not sure what would be the general structure though.

1. General Instructions
   • Project-agnostic guidance that applies to every project
   • Custom instructions for my local setup like commands (orbctl, deployment processes)
   • Practices that work well can graduate to global CLAUDE.md
2. Context / background info
   • System description relevant to the plan scope
   • High-level overview of current state:
     • Technical architecture
     • Key files and directories
     • Integration points
   • Foundation for early research
   • Link to related zettelkasten notes for deeper context
   • Brainstorming
     • Keep exploratory thinking in TaskPaper and link separately
3. User stories
   • Clear “As a user I want X so that Y” format
   • This approach surfaces misalignment between human intent and agent understanding
   • Creates shared vocabulary before technical implementation
   • Foundation for implementation design decisions
4. Implementation design
   • High-level architectural steps (following Kiro’s format)
   • Step-by-step approach that breaks into manageable development tasks

The plan file could be converted from a Markdown template into a Bike template 2025-07-20_13-14.