Great article! I am curious about the several instruction files, where you have one for each subproject. Do you have to include them manually when starting to work on a given subproject? Or is there a way to automatically include all of them at the start of a new session?
The point is to actually not include most of the instructions for every conversation since that bloats the agent context -> worse performance. Put things you need in every convo in CLAUDE.md. Other instruction files are supposed to be discovered by the agent when it plans the subproject task. So your project can look like this:
Great article! I am curious about the several instruction files, where you have one for each subproject. Do you have to include them manually when starting to work on a given subproject? Or is there a way to automatically include all of them at the start of a new session?
The point is to actually not include most of the instructions for every conversation since that bloats the agent context -> worse performance. Put things you need in every convo in CLAUDE.md. Other instruction files are supposed to be discovered by the agent when it plans the subproject task. So your project can look like this:
my-project/
├── CLAUDE.md # "See backend/BACKEND.md for APIs , infra/INFRA.md for deployment"
├── backend/
│ └── BACKEND.md
└── infra/
└── INFRA.md
Solid post. The CLAUDE.md split into BACKEND.md and INFRA.md is smart; I do something similar in OpenCode where you can layer custom instructions per project. The cost tracking point is real too. At $200-400/month you want to be deliberate about which provider you're routing through. Covered the config side of that in a guide I put together: https://reading.sh/the-definitive-guide-to-opencode-from-first-install-to-production-workflows-aae1e95855fb