LLM drafting versus template merge: separate flexible prose from fixed fields
Compare new prose from material with filling fields in fixed text, checking missing values, wording variation and batch revisions.
Choose an approach
Use template merge for fixed structures with confirmed fields and model drafting for explanation or rewriting. Combine deterministic fields with open prose. Models must not invent missing dates and templates do not verify facts. Keep sources and check fixed values and generated claims separately; a produced file is not verified content.
Variable substitution and new prose have different risks
Jinja renders variables and control structures; LLMs generate expression from prompts and material. Templates can leave blanks or fail on missing variables, while models may hide gaps with fluent prose. Fixed names and dates suit inspectable templates; explanations of differences can be generated separately. Define variable parts so style changes do not rewrite amounts or dates. Templates themselves are not factual sources.
Check missing fields and added claims against one source
Use complete and date-missing public material for both outputs. Check explicit missing values, invented commitments and final fields. Inspect different batch branches rather than only the first success. Generated factual paragraphs need source support. No real notice or message was created or sent here and no productivity result is reported.
Version template rules separately from generation requirements
Test variables and escaping after template changes and replay failures after prompt changes. Preserve input and version information to locate errors. Count design, review and rework; repeated fixed content need not be regenerated. Combine a fixed skeleton with checked open paragraphs and make direct fields distinguishable from interpretations for maintainers.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.