diff --git a/template/{{ part_dir }}/{% if layout == 'root' %}README.md{% endif %}.jinja b/template/{{ part_dir }}/{% if layout == 'root' %}README.md{% endif %}.jinja
index b946b8c..b23cc2f 100644
--- a/template/{{ part_dir }}/{% if layout == 'root' %}README.md{% endif %}.jinja	
+++ b/template/{{ part_dir }}/{% if layout == 'root' %}README.md{% endif %}.jinja	
@@ -17,4 +17,6 @@ make recreate
 
 {% if backend_services %}To deploy, `git pull && make deploy`. It runs `make pre-deploy`, which builds the image only when `pyproject.toml`, `uv.lock` or `Dockerfile` changed{% if database == 'postgres' %} and applies migrations{% endif %}, then `docker compose up -d`, then `make post-deploy`, which restarts the services whose mounted code changed. A deploy system with its own `up` step calls the two hooks around it. What was built and started is remembered as `refs/deploy/*` in the clone.
 
+In the fleet, production is deployed from `infra/komodo` instead (git mode of Komodo, see `docs/architecture.md` there): Komodo clones this repo, and a layer plus `deploy.sh` in the stack's directory tag images by the hash of what they bake in, write `.env` from sops and run migrations before the code. The repo only has to stay a plain recipe: images built from `build:` in compose, migrations as the `migrator` service when there is a database, `caddy/site.caddy`. `make deploy` and `refs/deploy/*` are for a checkout deployed by hand.
+
 {% endif %}`make fmt` formats, `make check` runs ruff and ty.{% if targets %} Also {{ targets | join(", ") }}.{% endif %}