Build from a whole document

Both a glossary and a style guide can be built from a document you have already worked with. You can point the build at just the page you have open, or at the whole document. The whole-document run reads every block, not a sample, so the terms it pins and the voice it captures reflect the full text, which is what you want before you reuse them across a project.

For a glossary, the Glossaries page’s New button splits into blank, Generate from document…, and import from a file. A card’s ⋮ menu has Auto-build from document…, which adds to that glossary rather than creating one. The editor’s Glossary tab has the same action, scoped to a single glossary once you have opened one.

For a style guide, the editor’s Style guide tab has Generate & review, and the Style guides page offers the same run under New → Generate from document….

Either way, pick the whole-document option to get the wave-by-wave read rather than a single page. Literess can take you to any of these screens and set one up with you, but she does not run the build herself.

Yes. The Languages section of both dialogs offers the document’s own target plus every language version it already has, with Add… for one it does not, and the estimate then reads as words × languages. Each language is a separate run with its own reviewable result, so the review walks them one at a time and closing it keeps the rest waiting. A language with no translated version yet still works, because the run reads the source text, but it can only learn that language’s conventions and not its voice, since there is no target writing to sample.

A long document does not fit in one pass, so the whole-document build splits it into waves. Each wave reads a stretch of the document and refines what the earlier waves found: a term the first wave was unsure of gets confirmed or dropped as later waves see it used again, and a style guide’s read of the voice sharpens as more of the text goes by. There is no block cap; a book-length document takes more waves.

A whole-document build is billed in words, the words it reads while drafting. That total is known before it runs, so you get an estimate first: the word cost as an “up to N words” ceiling, an estimated time, and the number of waves the document will take. You only start the run once you have seen it.

The ceiling is an upper bound, not a flat fee. A build that finishes early bills only the words it actually read, so the final charge is at or below the “up to” figure. Reusing translation memory or a glossary you already have costs nothing; only the reading is billed.

Nothing joins your library until you review it, and both kinds ask where the result should go before they start. A glossary build saves to a new glossary or to one you already have; a style guide build saves to a new guide or as a new version of an existing one. Nothing is overwritten either way.

A glossary build lands as a draft of candidate terms: accept, reject, or edit each one, and only the ones you keep become entries. A style guide build lands in the review described in Styleguides, an editable form for a new guide and an editable diff for an update, and nothing saves until you apply it.

A build that knows which glossary it is feeding reads that glossary before it reads a word of the document. Two things follow. It stops proposing terms you have already saved, so the review is what the document actually adds, not a list half of which is already saved. And where the document teaches an existing entry something it lacks, the run proposes an update to it: another spelling of the source word the text actually uses, a do-not-translate flag the usage implies, a rendering to add to the never-use list, or a sharper context hint.

The review keeps the two apart, under Adds N new terms and Updates N existing terms. Each update row spells out the current value against the proposed one and can be dropped on its own. Accepted updates add to the entry rather than replacing it: forms and never-use renderings join what is there, and a do-not-translate flag is never removed for you.

A translation you have recorded is left alone. The one exception is when the document shows it is wrong for how the term is used there, and then it arrives flagged as a replacement on its own row with a one-line reason, so you can decline the rewrite and keep the rest of that update. Updates ride the same run, so none of this costs anything extra.