I discovered that Claude Code, unlike current Agents in IDEs (Cascase in Windsurf, Cursor Agents), can follow instructions outlined in another file or in a large prompt. This allows scaffolding new architectures quickly and implementing specific applications on top.
- Since Claude Code will use the directory as context, I find it useful to hide context. This allows us to greenfield progressively. I find too much context results in undesirable results.
- Copy contents of current step (1-4) to
spec.md
in root of project - Prompt claude
follow the instructions in spec.md
- Verify output
- Remove spec.md from project
- Prompt Claude Code
/init
- Repeat until steps completed
python -c "import urllib.request; print(urllib.request.urlopen('https://gist.githubusercontent.com/buwilliams/db39f23acb1e6ad9595c99727448d530/raw/ae27d14504a83c0c6bbe831fce9a8907bda4f4c8/1-backend.md').read().decode('utf-8'))" | claude -p --allowedTools "Bash,Edit,Write"
python -c "import urllib.request; print(urllib.request.urlopen('https://gist.githubusercontent.com/buwilliams/db39f23acb1e6ad9595c99727448d530/raw/60e36dd48e38b6b8e5d1fc1d26e3fd8fe78b7b35/2-frontend.md').read().decode('utf-8'))" | claude -p --allowedTools "Bash,Edit,Write"
python -c "import urllib.request; print(urllib.request.urlopen('https://gist.githubusercontent.com/buwilliams/db39f23acb1e6ad9595c99727448d530/raw/60e36dd48e38b6b8e5d1fc1d26e3fd8fe78b7b35/3-design.md').read().decode('utf-8'))" | claude -p --allowedTools "Bash,Edit,Write"
python -c "import urllib.request; print(urllib.request.urlopen('https://gist.githubusercontent.com/buwilliams/db39f23acb1e6ad9595c99727448d530/raw/0e7387d0c896fd30e04f3ba8c91669226d51216e/4-application.md').read().decode('utf-8'))" | claude -p --allowedTools "Bash,Edit,Write"