Separating content (structured data) from presentation (templates), so one build step produces every version of a document from a single source.
Project requirement
Keep a CV and a cover letter in two languages and several variants without the versions drifting apart.
Technical question
How can one content source produce every language and variant, while missing translations and overlong documents are caught automatically?
Validation
A working generator builds two variants in two languages; tests cover resolution and escaping, and repeated builds give identical PDFs.
Decision
Resolve language and variant before rendering, keep templates free of logic, allow overrides only for whitelisted keys, and check the output during the build.
Resolution order: defaults, then variant, then per-document override, then command line.
Escaping depends on the output format (LaTeX or HTML), so it belongs to the rendering step, not to the content.
YAML data is resolved into a plain context in Python and rendered with Jinja2 into LaTeX, then compiled to PDF with latexmk.