The Supervertaler Editor

memoQ gives an add-in no window of its own, so everything Supervertaler needs a window for lives in one that runs alongside memoQ: the Supervertaler editor. Open it from the Start menu – it is called Supervertaler for memoQ there – or with Edit… in memoQ’s own Supervertaler options dialog.

It is where you choose what memoQ uses for each job (model, prompt, memory bank, termbases), edit the shared Supervertaler prompt library – the same folder of prompts the Trados plugin uses, so a prompt tuned in one tool is available in the other – draft prompts with AutoPrompt, describe a document’s figures with FigureLens, watch what the plugin is doing in the Activity window, connect AI assistants and manage your licence.

It also removes MT marks from documents you have exported, so that rows you reworked no longer say they came from machine translation.

Instructions can also be typed straight into memoQ’s settings dialog instead of coming from the library.

Choosing a prompt

In Resource console → MT settings → Supervertaler → Configure plugin, the Prompt dropdown lists every translation prompt in the library, grouped by folder. Pick one and its text appears (read-only) in the Instructions box below.

Choose (instructions below) instead to type your own; the box becomes editable.

The dropdown stores which prompt you chose, not its text. Edit the prompt anywhere – in the editor, in the Trados plugin, in a text editor – and memoQ uses the new version on the next segment.

Only prompts in the library’s Translate folder are offered. Proofreading and QuickLauncher prompts exist for other tasks and would produce commentary where a translation belongs.

The editor

Open Supervertaler for memoQ from the Start menu, or press Edit… beside the Prompt dropdown in memoQ’s options dialog.

• Left: the library as a tree, folders and prompts, and below them your Memory banks with the files in each – _shared first, marked always sent. Select a prompt or a bank file to open it.
• Right: name, description, which product it is for, sort order, and the prompt text with Markdown headings and {{PLACEHOLDERS}} highlighted.
• Toolbar, left: New, Save, Placeholder, AutoPrompt.
• Toolbar, right: whether Pre-translate goes to Claude Desktop, and the Activity window. Everything else is on the menus: File for the library, memoQ for the open job, Settings for Translation settings and Connect AI assistant, and Help for this documentation and your licence.

The Claude Desktop button on the right is a switch, not a command, and its caption says which mode is on rather than what pressing it would do. It decides what Pre-translate does with rows nothing was staged for – translate them through your API key (off) or leave them as they are for the chat (on) – so it is worth a glance before a long run. It is the same setting as Settings → Pre-translate via Claude Desktop, and as the checkbox in memoQ’s own dialog – change it anywhere and all three follow.

Save with Ctrl+S. A built-in prompt – the Default Translation Prompt, say – is never saved over: your changes are saved as a prompt of your own, named ”… (my version)” unless you gave it a new name, and it becomes the active prompt if the original was. The built-in one stays as it shipped, so an update to it can never overwrite your work.

Any settings the file carries that the editor does not have a field for – tags, favourites, QuickLauncher flags set by the Trados plugin – are preserved untouched on save. The status line under the description says which ones the file has.

A memory bank file opened from the tree – brief.md, terminology.md, style.md or any other article – is shown as plain text, without the prompt fields, and Ctrl+S saves it. Right-click a bank, or a file in it, for Set as active memory bank, which chooses it for the open project as the Bank row does. _shared cannot be chosen that way: it is always sent, whatever bank is chosen.

What memoQ is using

The bar under the toolbar opens with the memoQ project these apply to, because everything on it is recorded against a project. When it reads no project yet, in red, memoQ has not sent a translation request and the plugin does not know where it is – usually because Supervertaler is not selected as the MT engine in a newly created project. A memory bank chosen at that moment is filed against whichever project came before, silently, so it is worth a glance before changing anything.

Then four things memoQ will apply to every translation:

Model, Prompt and Bank. Click any of them to change it. Each opens a list with a filter box rather than a dropdown menu, because they grow with the work – a prompt library reaches forty entries quickly, and a bank per client does the same. The bank can also be set to nothing.

Termbases is the fourth row, and it is a summary rather than a picker: BRANTS (project) + BEIJER · 1 of 2 to the model – which termbases are ticked for this project, which of them is the project’s, and how many reach the model. A termbase selection is a set, and a dropdown cannot hold one; clicking the row opens the Termbases window, which can. It reads none in grey when terminology is off.

These are the choices that change between jobs, which is why they are here rather than in Translation settings: they are what the model knows before it is shown a segment. Each is also the same setting memoQ’s own dialog shows, so either place can change it.

The memory bank is remembered per project. Choose one while working on a job and it comes back when you return to that job – and a project you have never chosen one for does not inherit the last one, because a bank carries one client’s terminology and the wrong one is worse than none.

What such a project gets instead is the row called (no client bank – shared defaults only). Your _shared bank still travels: it is where the material that applies to every job regardless of client lives, so it is never switched off by not choosing a client. See Memory banks for where they live.

The shared bank is made for you (from v0.1.26). When memoQ loads Supervertaler, or the editor opens, and there is no _shared bank, it is made – in your team folder, if you use one – with a starting brief, terminology table and style guide. It is at the top of the memory banks in the editor’s tree, marked always sent – open its files there to fill it in – and it sends nothing to the model until you write something in it. A _shared bank you already have is left exactly as it is, including any file you deleted from it. It is not offered in the Bank list, because it is not a bank you choose; typing _shared as the name in New memory bank… there gives you the shared bank and leaves the project’s bank as it was. Before v0.1.26 the underscore was dropped, which made an ordinary bank called shared that is never sent as the shared one – see Troubleshooting if you have one. Supervertaler for Trados makes _shared the same way, from v18/19.20.201.

A bank is sent whole with every translation request, up to about 32,000 tokens, and to AutoPrompt up to 40,000. Anything that does not fit is dropped by priority and named in the Activity window rather than lost quietly – if you see a file listed there, that is the budget, not a fault. The cost of carrying it is small because the same text is sent every time and providers cache it: on one 370-segment run the bank and prompt together came to 45,870 tokens, and caching turned roughly $10 into roughly $4 – for Pre-translate and, since September 2026, for the single rows memoQ asks about as you move through the grid.

A large bank sends only what the job needs. Once your banks together pass about 8,000 tokens, each job gets a selection instead of everything: only the terminology rows whose term occurs in the document (in either column, whichever way the table was written), and the notes the model judges relevant after reading the document, asked once per job. The client bank’s brief, style guide and domain notes always go. The house rules in _shared are judged one by one in the same question: a rule whose Scope: line clearly does not fit the job’s language pair or kind of document is left out, so a rule written for Dutch-to-English work stays out of an English-to-Dutch job. When the model is unsure, the rule goes; if the question cannot be asked – memoQ has not yet shown enough of the document, or the request fails – everything goes, as before. It happens by itself; there is nothing to set. The Bank row’s tooltip says when it is happening and names a file under C:\Users\<you>\Supervertaler\memoq\bank-extracts\ listing what was sent and what was left out, which is the place to look if a translation seems to have missed something you wrote down.

Notes for the assistants only. Some notes are about how to work a job – what to check before delivery, how to verify a termbase – and help an AI assistant rather than a sentence being translated. Put this at the very top of such a note:

---
audience: assistant
---

It is then left out of every translation request (the translation engine and AutoPrompt), while Claude Desktop and ChatGPT still read it through the Supervertaler tools. A note without it is sent as always, and the job’s selection file lists anything left out this way.

You can switch the bank off altogether: Settings → Translation settings → Send the memory bank to the model. Off, nothing from any bank – the client’s or _shared – goes to the model, for translation or AutoPrompt, and the Bank row reads not sent. The bank chosen for each project is kept for when you switch it back on.

Placeholders

Prompts use {{SOURCE_LANGUAGE}} and {{TARGET_LANGUAGE}} rather than naming languages, so one prompt serves every language pair. memoQ fills them in per project.

Insert placeholder lists the ones memoQ can fill. A placeholder memoQ cannot fill – {{SOURCE_SEGMENT}}, say, which only the Trados plugin provides – is shown in red, and the editor warns that it will reach the model as empty text. Do not use those in a prompt meant for memoQ.

Which product a prompt is for

Available in can be both, trados or memoq. memoQ’s dropdown hides prompts marked for Trados only. Leave it on both unless a prompt genuinely depends on something one product cannot supply.

A prompt tied to one product says so in its filename: Patent claims EN-NL [memoQ].md, Define [Trados].md. Prompts available to both carry no marker, so the absence of one is itself readable – which is the point, because in Explorer the metadata header is not visible and every prompt otherwise looks alike.

The marker is written from the Available in field on every save and stripped again on every read, so it is a label rather than a setting. Renaming the file in Explorer does not change which product a prompt is for, and the next save puts the old marker back: change the field, not the filename. The editor’s tree and the Prompt dropdown show the same thing in words.

Where the library lives

C:\Users\<you>\Supervertaler\prompt_library\ – one Markdown file per prompt, with a small metadata header. Open folder in the editor takes you there. The files are plain text; nothing stops you editing them directly, and a folder synced between machines carries the whole library with it.

AutoPrompt: drafting a prompt for the open project

Press AutoPrompt… in the editor’s toolbar, or choose it from the memoQ menu. Supervertaler reads the document you are translating, the termbase hits in it (from termbases ticked AI) and anything you have already confirmed, and has the AI write a prompt tailored to that job – domain, register, a locked glossary, the lot. The result is saved under Translate and opened for you to review; then pick it from memoQ’s Prompt dropdown.

Before it runs you choose the document (if several are captured), and can add a briefing – client, audience, style, what to avoid – which the AI treats as authoritative. The briefing is the one input nothing else supplies: the filing route, a discrepancy you already know about, anything true of this job that is not in the document or the memory bank.

The two checkboxes grey out when they have nothing to offer – no termbase is ticked AI for this project, or nothing has been confirmed in this document yet – and say which it is, rather than sitting there ticked and doing nothing.

Preview context… shows you exactly what will be sent, before anything is sent: the extract from your document, the termbase hits, the segments you have confirmed, the briefing you typed, and the instructions the AI is given about writing a prompt for memoQ. It makes no API call and costs nothing, and the briefing box stays open behind it – so the loop is look, add what is missing, look again, then generate.

Three things to know:

• memoQ must be running with a Supervertaler engine active, and the document must have been captured – one Pre-translate does it (free, with the Claude Desktop box ticked). The plugin only sees what memoQ has sent it.
• It uses the provider, model and API key from your Supervertaler settings. Two calls: a short one to classify the document, then a long one to write the prompt. Expect a minute or two.
• The prompt is written for memoQ, not copied from the Trados recipe. Single-segment lookups are handled as well as batches; tag markers must be reproduced exactly; translations you have confirmed outrank the prompt’s own glossary; and it is kept to 1,500–3,000 words because memoQ re-sends the whole prompt with every ten-segment request.

Draft it again later in the job and it gets better: by then it can see what you have confirmed, which is stronger evidence of how you want this document translated than the source text alone.

A drafted prompt is the only source of terminology

A prompt AutoPrompt wrote ends in a locked-terms table chosen for this document. So while one is selected, the termbases’ preferred renderings are not sent to the model as well – two lists of terminology that were never written to agree, with nothing saying which wins, is a worse position than one list.

Forbidden terms still go. A preferred rendering is advice, and two sources of advice can contradict each other confusingly; “never use this word” is a constraint, and there are few of them. So they travel whatever prompt is selected.

Nothing else about the termbases changes: they still drive the terminology pane, the QA check and AutoPrompt’s own reading of the document. Only the per-request injection stops, and the Activity window says so once per prompt.

The consequence worth remembering: a term you forbid after a prompt was drafted is enforced immediately, but a preferred rendering you add afterwards is not – draft the prompt again to take it in. Termbase from this prompt’s terms is the other half of that loop: derive the project termbase from the prompt and the two cannot contradict each other in the first place.

Prompts saved from the chat over MCP count as drafted too, and are marked for the product you were connected to.

List numbering reaches the model as structure

The letters on the steps of a claim – a), b), c) – are not text. Word generates them from the paragraph’s list settings, so memoQ’s grid does not contain them and neither does anything the plugin is sent. Shown six unlabelled sentences followed by steps a. to f., a model will flag the reference as a possible source defect, and it will do so on every lettered list in every document.

With the live document link connected, the plugin knows which file memoQ imported, reads the numbering out of it – counted over the whole document, exactly as Word renders it, restarts and all – and sends each paragraph’s marker in front of its first segment as [#e)]. The prompt tells the model this is structure: use it to resolve cross-references and keep list items parallel, never translate it, never reproduce it. Every reply is checked before it reaches the document, and an echoed marker is removed, with a line in the Activity window saying so – that line is your evidence, per model, that the rule is being obeyed.

There is no switch for it. Without the live link, or for a document with no lists, the model is instead told that numbering is supplied by the document and not to flag its absence. The Activity window says which of the two happened, once per document. On a project checked out from a server the file memoQ names is on the project manager’s machine, not yours; locate it once in FigureLens and the numbering is read from your copy.

Translator comments

Where a note is genuinely necessary – an ambiguity in the source, a term that could go two ways, a probable defect in the original, an error it corrected in a 100% translation memory match, a deliberate departure from the terminology – the AI puts it inline at the end of the target as a [[TC: …]] marker: one per segment at most, short, and written as your comment to the client. This is built in, not something a prompt has to ask for: every request ends with a fixed output contract that says so, whichever prompt is selected. A prompt can add to it – its own examples of good comments, a language other than English for them – but cannot switch it off by accident. Supervertaler for Trados uses the same form, so a prompt written for one product reads correctly in the other.

What the model may send back

The same contract says the reply is the translation and nothing else, and every reply is checked before it reaches the grid. One that carries commentary anywhere but a single trailing [[TC: …]] – a note after the translation, the model explaining its own choices, markdown or a line break the source does not have, a second marker – is asked for once more with the contract restated. If the second reply does the same, nothing is written: the row is left for you, and memoQ shows under it why, together with what the model sent, so you can still use the translation part if it was sound. Each retry and refusal appears in the Activity window.

Tags that differ from the source are reported there too, but never refused: a translation that drops a formatting tag is often right, and a row left empty for it would cost more than it saves.

A row whose source has nothing to translate – only a placeholder tag, spaces or punctuation – is not sent to the model at all. Its source is copied into the target, tags intact.

Nothing extracts these for you, and that is deliberate. You read them in the grid as you review, decide which are worth keeping, turn those into real memoQ comments on the segment, and delete the marker from the text. Search for [[TC: to find them all.

FigureLens

The model sees a document’s text and not its pictures. A claim that names part 12 is translated by a model that has never seen part 12, and a figure’s caption is often the only description of it anywhere in the text. FigureLens… on the toolbar (also memoQ → FigureLens…) is the panel that closes that gap, in two steps. It is named for what it does beside TermLens: that one shows the model the terms in a segment, this one shows it the pictures.

Step 1 – Extract images copies every image out of the documents into the active memory bank’s figures\ folder, named after their figure numbers – Figure 01.png, Figure 02.png – so a folder of drawings reads like the document. Free, no AI. The panel says how the labels were arrived at: paired by position and checked, taken from nearby text, or withheld when it could not tell.

Step 2 – Describe images with AI shows each image to the model, together with what the text says about it, and saves the descriptions as figures.md in the memory bank – one paid request per image, and the panel states the count and the provider before you click. Every prompt reads that file from then on, so read it first: a wrong caption would be invisible and everywhere. Reference signs the model reads in a drawing that appear nowhere in the text are listed at the end, because that is a defect worth raising with the client before filing. Describe from the text only is the free alternative – what the document itself says about each figure, without looking at the images – and either replaces the other, after asking.

The images folder is not chosen: it is inside the memory bank, because that is where figures.md goes and the two belong together. With the shared bank or no bank active, the Result line offers to create a bank named after the memoQ project and switch to it.

Where the documents come from

memoQ never keeps the original file in its project folder – a local project stores the filename as an empty placeholder, a project checked out from a server stores only memoQ’s own data – so the panel works from the file memoQ imported, wherever that was:

• On a local project, the live document link reports the path memoQ imported each document from, and the panel finds the file there without any setting.
• On a server project, memoQ records the path the file had on the project manager’s machine, which does not exist on yours. The panel lists those documents as not on this computer and offers Locate the original document… – point it at the copy you were sent, once, and it is remembered for that document in C:\Users\<you>\AppData\Local\Supervertaler.memoQ\document-files.txt. Locating a document here also switches on list numbering for it.
• Add a document file… is for a Word file memoQ has said nothing about at all. Its images are read from the file directly; the file is remembered for the active memory bank.

Supervertaler finds the project folder by asking memoQ where it keeps its projects – the custom folder set under Options → Locations → Projects, the default C:\Users\<you>\Documents\My memoQ projects when you have not set one, and memoQ’s own register of every project, which names each one’s actual folder. Moving your projects folder therefore needs nothing here, and projects left behind in the old location are still found.

Document images report at the bottom writes a Markdown listing of every image in every document – label, size, caption, the text around it – into the memory bank and opens it. No AI call.

Termbase from this prompt’s terms

An AutoPrompt draft ends with a locked-terms table – a dozen or so renderings chosen for this document. That table is exactly what the terminology plugin and the check_terminology QA tool should work from: a general termbase flags application → aanvrage in every paragraph of a software text, a project termbase knows better.

Choose memoQ → Termbase from this prompt’s terms… with the prompt open. Supervertaler reads every table in it that names a source and a target column, turns notes of the form never “apparatus” into forbidden entries, and makes a termbase named after the prompt – or adds to it if one exists, skipping pairs already there – and ticks it Read and Project for the current memoQ project. A row that lists alternates, regenereren / regeneratie | regenerate / regeneration, becomes one pair per alternate. It works whether or not memoQ is running; memoQ picks the change up within a few seconds.

Any prompt with a table laid out the same way works, not only AutoPrompt’s. Correcting the result afterwards is what the Terms window is for; rows the prompt no longer has are not removed, since they may have been corrected by hand.

Settings

Settings → Translation settings holds how Supervertaler translates: provider, model, endpoint, parallel requests, segments per request, and whether termbase hits, surrounding segments and the memory bank are sent to the model. These are the same settings as memoQ’s own Supervertaler dialog, reading and writing the same file, so either place can change them and both show the same values.

The Model list is short on purpose: three to five models per provider, each with a line saying what it is for. A provider’s own catalogue runs to thirty or forty entries – image models, speech models, dated snapshots of the same model – and a list like that is one nobody in a hurry can choose from. A model that has been superseded is removed rather than annotated, so what is left is what is worth using today.

Fetch list asks the provider for its full list, using the API key below. Show all models then shows everything it returned under the short list, which is where to look for a model released after your copy of Supervertaler was built. The fetched list is remembered, and so is the tick, so this is a decision you make once. The line under the button says when the list was last fetched and how much of it is beyond the short list.

The box stays typeable throughout, so a gateway, a private deployment or a model that appears in neither list can be entered by hand – and a model already saved in your settings keeps working whether or not it is in the list on screen.

Changing the provider changes three things together: the model list, the model itself – to that provider’s first recommendation, since a model belonging to another provider can only fail – and the API key, which is re-read for the provider you have just chosen. A key you type here is remembered per provider while the window is open.

memoQ’s own Configure plugin dialog has the same three controls, reading and writing the same settings, so it does not matter which one you use.

Segments per batch can only lower what memoQ does, not raise it – memoQ hands a plugin about ten segments at a time during Pre-translate, however high this is set. Lowering it is still worth doing if a model keeps returning fewer translations than it was sent.

Send anonymous usage statistics when memoQ starts (from version 0.1.1) is your answer to the question the editor asks the first time it opens. If it is ticked, memoQ sends one small message each time it starts: a random ID, the plugin version, the Windows and memoQ versions and your system language – never your documents, translations, terminology or file names. Nothing is sent unless you tick it. See the privacy policy.

Settings → Pre-translate via Claude Desktop (MCP) is on the menu itself, and on the right of the toolbar, because it is the one that gets flipped between jobs rather than set once. See MCP server.

API keys live in one file

Every Supervertaler product reads the same file:

C:\Users\<you>\Supervertaler\settings\api-keys.json

One key per provider, plain text, editable in Notepad. A key pasted here works in Supervertaler for Trados and Supervertaler Sidekick as well, and rotating one means changing one line in one place. Before this file existed there were three dialogs in three products each keeping their own, which is how an hour goes missing to a key for one service pasted into another’s box.

Keys are stored under the provider ids claude, openai and gemini. Sidekick keeps its machine-translation keys in the same file under their own names – note that google there is Google Translate, not Gemini.

Plain text is deliberate, and the same choice Supervertaler for Trados has always made: a key that can be rotated by pasting a line into a text file is a key that actually gets rotated, and anyone who can read that file can already read everything else in your profile.

The API key box shows the key for the provider you have selected and writes back to that file. If you had a key configured before the file existed – in memoQ’s own settings, or in Trados’s – it is copied in the first time Supervertaler needs it, so there is nothing to do.

If you paste a key that plainly belongs to another service, the line under the box says so as you type: This is an OpenAI key, not an Anthropic one. That is worth more than the provider’s own answer, which is that the key is incorrect.

The Activity window

memoQ’s Pre-translate dialog is modal and says only Processing, for as long as the run takes: no engine, no model, no count, and no sign when something is wrong. memoQ → Activity…, or Ctrl+L, opens a window that shows what Supervertaler is actually doing.

It is a window of its own rather than a panel so that it can sit over memoQ while that dialog holds the screen. Tick Keep on top and you can watch a Pre-translate run from the first batch to the last.

What it shows: the engine and model each project starts with, the termbases as they load and how many terms came out of them, a warning when the selected prompt faces the opposite language pair, every batch with the segments sent, the segments returned and the terms matched, AutoPrompt drafts, and anything that failed. A batch that comes back short is called out rather than logged flatly, because that is the failure that quietly shifts every translation after it.

Five lines are worth knowing by sight:

• Bank – which memory bank a project switched to, and once per job how much of it is being sent. If it ends with a file listed as not sent, that is the budget trimming by priority, not a fault.
• Terminology – said once when a drafted prompt is holding the termbases’ preferred renderings back, so a quiet change to what reaches the model is never silent.
• Rows – after each Pre-translate call, where its rows came from: how many were staged by an assistant, how many the model wrote, how many were copied because there was nothing to translate, and how many were left for you. Staged and model-written rows look the same in the grid; this is where the difference shows, and the totals line adds them up over the run.
• Reply – a reply that broke the output contract: asked for again, refused, or served with tags that differ from the source.
• The token count on each batch – tokens: in 1,041 (cache write 45,870) out 1,233. The prompt and the bank are identical on every request of a run, so providers cache them: the first batch writes, the rest read at a tenth of the rate. If cached never appears across a long run, something is re-sending the block at full price.

Show everything un-hides the per-request diagnostics – memoQ’s capability probes, lookup sessions, single-segment translations – which are what you want when something is wrong and noise the rest of the time.

The window reads the plugin’s own log, C:\Users\<you>\AppData\Local\Supervertaler.memoQ\plugin.log, rather than being fed by the plugin. So it shows what happened before you opened it, it works whether or not memoQ is running, and closing it costs nothing. Its position and size are remembered.

Drafting prompts with Claude

If you use the MCP server, Claude can write into this library: “Draft a translation prompt for this project and save it.” It reads the captured document, your confirmed segments and your termbases, saves the result as a new prompt, and you pick it from the dropdown. The editor is where you review and tune what it wrote.