doc-templates

Document templating

Using render_template

Call the render_template tool with:

• template — the template text, containing ${path.to.var} placeholders (same ${...} syntax already used elsewhere in PawFlow for flow parameters and secret references — nothing new to learn).
• variables — a nested JSON object resolved by dotted path, e.g. {"dossier": {"client": "...", "type_affaire": "..."}, "delai": {"date_butoir": "..."}}.

The tool never guesses a missing value: any placeholder whose path is not found in variables is left as an explicit [path manquant] marker in the rendered text, and the tool's result also lists every missing path separately (missing_variables). Always surface unresolved placeholders to the user before presenting the draft as ready — a document with a silent gap is worse than one that visibly flags it.

The tool only transforms text; it never writes a file itself. Present the rendered draft to the user, or write it yourself with your own tool grants if asked to save it — always as a draft, never as a sent/filed document without explicit human validation.

Bundled starter templates

Three starter templates ship bundled with this skill, under templates/ (sibling files of this SKILL.md, mounted alongside it wherever the skill is mounted): mise_en_demeure.md, accuse_reception.md, convocation_rdv.md. These are examples to copy and adapt, not fixed content — a firm using this package should customize them (letterhead, standard clauses, tone) rather than send them as-is. Treat the shipped copies as a starting point that becomes the firm's own editable version after first use, never re-render the signed package content directly for every use if the firm has since edited its own copy.

Importing an existing .docx/.pdf as a starting point

When a user wants to turn an existing document into a template:

• Content-only import (a .pdf, or a .docx where only the text matters): extract the text, drop it into a new template draft, then the user marks variable spots themselves by typing ${...} placeholders into the extracted text.
• Formatting-preserving import (a .docx with letterhead/layout worth keeping): keep the original .docx as the template file; placeholders are inserted directly into its text runs, and rendering substitutes those runs in place rather than rebuilding the document from markdown. Never target a .pdf for formatting-preserving rendering — a PDF has no reliable editable structure for in-place substitution.
• Suggesting variable spots is assisted, never automatic. You may propose likely variable locations (names, dates, amounts detected in imported text) as highlighted suggestions to validate one by one — never insert them directly. Guessing that a name is a variable is not different from guessing a variable's value: both remain a proposal for the user to validate, never an applied fact.