Styleguides

Fill in what matters and leave the rest:

• Fields — short dropdowns for tone, register, audience, address form (Sie/vous vs. du/tu), sentence rhythm, localization, quote marks, number and date format.
• Reference samples — passages of source or target text used as voice references. A target sample (“here’s what good output sounds like”) is the strongest signal of all, especially for fiction and marketing.
• Rules — short do/don’t bullets (“Always use Oxford commas.” “Don’t translate brand names.”).

Begin with a preset (literary modern, marketing punchy, technical precise) and tweak, or paste a passage of your existing translated work and let Literess fill in the fields for you. Each edit creates a new version and activates it, with older versions kept for rollback.

Generation never lands on its own. The editor’s Style guide tab has a Generate & review button, and the Style guides page starts the same run from New → Generate from document…. It drafts a guide from the page you have open or from the whole document, and before it starts you choose where the result should go: a new style guide, or an update to one you already have. Either way the result opens in a review first.

A new guide opens as one editable form. Remove a field, reword a rule, drop a sample, rewrite the notes, then save. An update opens as a diff against the guide’s current version, with what the update adds highlighted and what it would drop struck through, and every change editable, rejectable, or restorable on its own. The footer counts what Apply will actually write, and applying creates a new version of the target. Nothing is overwritten.

The draft arrives with all ten style settings filled, not only the two or three the document happened to evidence, so a setting with nothing behind it comes back at its neutral default and never sits blank. Generation also reads the glossaries switched on for that document, so the draft neither restates a term you have already decided nor contradicts it. Vocabulary shows up as a pattern (the register the words sit in, how foreign words are handled), never as term-level rules, which belong in the glossary.

If you close the review without deciding, the result waits. A Review pending badge sits on the guide’s card in the library and in the editor’s Style guide tab, and the same badge appears at the top of the guide’s own page with a Review button next to it. It survives a reload; it is stored, not held in the browser. That page also carries the library card’s other actions behind a ⋮ button: duplicate, merge, move to a team, and delete.

Words are billed when the generation runs, so a result you discard still costs what the run read. The dialog says so before you start.

One guide can hold a separate, complete version for each language you translate into. Above the fields, samples and rules sits a language strip: Language-agnostic first, then one tab per language you add with +. Each tab edits the same three sections as a full guide in its own right.

Nothing inherits and nothing merges between them. A run translating into German uses the German version exactly as it stands; a language with no version of its own uses the language-agnostic one. That is the whole rule, and it means a version you leave half-filled stays half-filled rather than quietly borrowing from somewhere else.

Reference samples are source→target pairs: the source passage and its translation. The target side can stay empty until the passage is translated, and the source still informs runs in the meantime. Notes are guide-wide rather than per language. Language versions live in the draft until you press Save changes, which writes them as a new version like any other edit; removing one asks for a plain confirmation first.

When an active guide has no version for the document you are working on, its card in the editor sidebar says so: Active · language not assigned, or Active · shared only; derived for German. Next to it is a Generate {language} version button that opens the generation dialog already pointed at this guide and this document.

The result is filed by kind. Address form, quotation marks, number and date formats, and target-language samples go into that language’s version; the voice rules stay in the shared, language-agnostic part. The review tells you which is which before you apply anything, and every run that saves into an existing guide is routed this way, so a language’s conventions can never end up rewritten as a shared rule that applies to all of them.

If the guide carries no language label of its own, Transept first works out which language its current content is in, from its samples and rules, and files that content under that language. Where it cannot tell, it assumes the document’s language and says so in the review (“Existing content → German · Generated content → Ukrainian”), which you can correct before applying. Reference samples come from real work wherever possible: if the document already has translations, sample-sized passages become source→target pairs rather than invented examples.

Generating into a guide never disturbs the versions for other languages, and nothing changes until Apply creates and activates one new version.

The style guide card in the editor sidebar carries a chip per language version. Selecting one shows that language’s resolved guidance read-only, with a note when the version you are looking at does not apply to the document in front of you. This doc returns you to what runs actually read, and the pencil opens the full page for edits. The card’s status line names the version in play: Active · Ukrainian version applies.

Merge sits where it does for a glossary: the ⋮ menu on the library card, the same menu on the guide’s own page, and the editor sidebar. Pick the guide to bring in. Only the guide you merge into changes, and the merge costs nothing.

A style guide merge is an AI step. It rewrites the two guides into one coherent guide instead of stacking their rules together: the guide you merge into wins a direct conflict, near-duplicate rules collapse into one, and per-language sections and reference samples from both sides are kept. It shows its reasoning while it works and you can cancel it, the same as generating a guide.

It saves nothing by itself. The merged result opens in the review screen a generated update opens, where you edit it and apply it as a new version, or discard it. Style guides never need languages assigned to merge.

Literess can read the merge screen and change its settings while it is open, and the same goes for a generation screen. The button that starts the work stays yours. See What Literess can do.