AutoPrompt

AutoPrompt uses AI to analyse your entire project and generate a comprehensive, domain-specific translation prompt tailored to your document. The generated prompt includes terminology rules, style guidelines, anti-truncation controls, and domain-specific instructions – ready to use with Batch Translate.

How It Works

1. Start the analysis

On the Batch Operations tab, click the AutoPrompt… link next to the prompt dropdown.

2. What gets analysed

Supervertaler gathers the following data from your project and sends it to your configured AI provider:

Data	Purpose
All source segments	Domain detection, document content analysis, project context
Termbase terms	Filtered to only document-relevant terms (TermScan), then included as a locked glossary in the generated prompt
Translated segments	Human-confirmed segments only (Translated, Approved, or Signed-off status) – used as TM reference pairs and style anchors. Unconfirmed AI-generated translations are excluded.
Language pair	Embedded in the generated prompt

Note

The full document is sent to the AI for analysis. For a typical 30,000-word document, this costs approximately $0.20–$0.25 with a Sonnet-class model, or $1.00–$1.15 with an Opus-class model.

2b. TermScan – automatic glossary extraction

Before building the prompt, AutoPrompt runs TermScan: it concatenates all source segments in the document and checks each termbase entry against this text. Only terms whose source term, source abbreviation, or source synonyms actually appear in the document are included in the generated prompt.

This dramatically reduces the glossary size – for example, a general patent termbase with 2,680 entries might yield only 123 relevant terms for a specific document. The status message in the AI Assistant confirms the filter: “Termbase terms (filtered 123 relevant from 2,680 total)”.

The filtering is case-insensitive and checks all variants of each term (source term, abbreviation forms, and synonyms). Terms that do not appear anywhere in the source text are excluded entirely.

Caution

TermScan filters by content, not by domain. A patent termbase contains many common technical words – “system”, “board”, “fan”, “screen”, “installation” – that will match nearly any engineering document. TermScan will include those entries even though they belong to a completely different translation context, and the AI will be forced to follow them. Before running AutoPrompt, disable every termbase that does not belong to the current project’s domain. Use AI Settings → Termbases included in AI prompts to control which termbases contribute to the generated prompt without affecting your TermLens display.

Caution

Termbase quality matters. Only enable termbases in AI Settings if you are confident they contain accurate, high-quality terminology for your project. A poorly maintained termbase with incorrect or outdated translations will constrain the AI and produce worse results. Modern LLMs – especially Opus-class and GPT-4-class models – are often better at choosing the right translation on their own than when forced to follow a low-quality glossary. When in doubt, disable termbases and let the AI translate freely, then add terms incrementally as you review.

3. Domain detection

Before sending to the AI, AutoPrompt runs a local keyword-based analysis to detect the document’s domain. Supported domains:

• Patent – claims, embodiments, prior art, figure references
• Legal – contracts, clauses, statutory references
• Medical – clinical terms, dosages, ICD/ATC codes
• Technical – specifications, software terms, standards
• Financial – figures, IFRS/GAAP, regulatory language
• Marketing – brand, audience, campaign language
• General – fallback for mixed or unclassified content

The detected domain determines which template the AI uses to generate the prompt – including domain-specific roles, rules, and section structure.

4. Review and refine in the AI Assistant

The generated prompt appears as a message in the AI Assistant chat. You can:

• Read through the prompt to verify it matches your project
• Ask follow-up questions to refine specific sections (e.g., “Make the glossary section more strict” or “Add a rule about chemical formula formatting”)
• Iterate as many times as needed – each refinement builds on the conversation history

5. Save the prompt

When you are satisfied with the generated prompt:

1. Right-click the assistant message containing the prompt
2. Select Save as Prompt…
3. Enter a name for your prompt (e.g., “DLCH Patent NL-EN”)
4. Click Save

The prompt is saved to the Translate category in the Prompt Manager and immediately appears in the prompt dropdown on the Batch Operations tab.

Note

Generated prompts are formatted in proper Markdown – ## headings for each major section, - bullet lists, **bold** for emphasised terms, and a Markdown table for the project-specific glossary. Open one in Obsidian, VS Code, GitHub, or any Markdown-aware viewer and it renders cleanly with a navigable outline. The prompt is also still a perfectly valid system prompt for the translator AI – the Markdown markup is structural, not output instruction.

Translator’s Comment methodology (always-on)

Since v4.19.111, every AutoPrompt-generated prompt embeds the Translator’s Comment (TC) methodology by default, regardless of source language or domain. The methodology asks the translator AI to silently correct obvious mechanical defects in the source (typos, broken words, hanging mid-sentence breaks, doubled spaces, stray punctuation, reference-numeral mismatches that are unambiguous in context, missing diacritics, etc.) and append a single concise comment at the end of the segment in this exact format:

⟦TC: short factual description of the fix(es)⟧

• The brackets are the mathematical white square brackets U+27E6 (⟦) and U+27E7 (⟧). These characters do not occur in source documents, so they are safe as out-of-band markers that can be extracted reliably in post-processing.
• One marker per segment maximum; multiple fixes are joined with semicolons inside one marker.
• Segments with no defects emit no marker.
• When the translator AI inserts a word or short phrase to fill a clear gap, that supplied text is wrapped in standard ASCII square brackets [like this] inside the running translation, and the trailing marker references it (e.g. ⟦TC: [bracketed text] supplied to close hanging sentence⟧).
• Numerical values, dates, dosages, claim language, statutory references, headings, identifiers, and proper names are never silently “corrected” – defects in those zones are preserved verbatim, with an optional ⟦TC: source ambiguous – ...⟧ marker if doubt exists.

The defect categories that count as “obvious” are adapted to the actual source language by the LLM (Dutch -d/-t verb typos, German missing umlauts, French accent slips, Spanish/Italian conjugation typos, etc.).

Note

The markers appear inline in the target segment in Trados Studio – they are not yet auto-extracted into the Studio comments panel. Extraction into the existing Studio comment infrastructure is a planned follow-up; the spec is locked (⟦ and ⟧ delimiters never collide with source text) so the extraction step is small once it gets prioritised.

Caution

Want a generated prompt without the TC methodology? Edit the generated prompt in the Prompt Manager after creation and remove the TRANSLATOR COMMENT FORMAT section plus any TRANSLATION MANDATE language about silent correction. A per-project opt-out via a UI toggle may be added in a future version – open an issue if you’d like to see it.

What the Generated Prompt Contains

A generated prompt follows the structure of professional translation prompts used by experienced translators. Depending on the domain, it typically includes:

• Role – domain-specific translator role with expertise areas
• Translation mandate – strict rules against simplification, paraphrasing, or “improving” the source
• Anti-truncation controls – explicit prohibition of omitting repetitive phrases or collapsing clauses
• Input handling rules – instructions for segment-by-segment translation in Supervertaler
• Domain-specific style rules – mandatory term mappings, register requirements, formatting rules
• Terminology hierarchy – priority order: TM matches > project glossary > domain conventions
• Preflight self-check – internal verification step before producing output
• Post-translation integrity assertion – completeness and faithfulness check
• Project context – AI-generated summary of what the document is about
• Project-specific glossary – all termbase terms, marked as locked and mandatory
• TM reference translations – validated translation pairs as style anchors
• Output format – translation only, no commentary, preserve formatting

Tips

Start with a confirmed translated sample

The generator includes confirmed segments (Translated, Approved, or Signed-off status) as reference pairs – up to 50, sampled evenly across the document. This gives the AI concrete examples of your preferred style and terminology, resulting in a more accurate prompt. Unconfirmed segments (e.g. from a previous AI batch translation that you haven’t reviewed yet) are excluded to avoid feeding unverified output back as “correct” references.

Tip: Before generating a prompt, confirm a handful of segments you are happy with. Even 10–20 confirmed segments give the AI meaningful style anchors to work from.

Only enable termbases that belong to this project

Before clicking AutoPrompt, go to AI Settings → Termbases included in AI prompts and enable only termbases that are directly relevant to the current project. Disable everything else – including large general-purpose termbases, termbases from other clients or domains, and any termbase you are not actively maintaining for this project.

This setting is independent of your TermLens display: disabling a termbase for AI context does not hide its chips in the editor. You can keep a termbase visible for reference while excluding it from the generated prompt.

Note

Prefer compact, project-specific termbases. A small termbase of 50–200 carefully curated entries for this client will produce a far better glossary than a general termbase of 2,000 entries, even after TermScan filtering. Large termbases increase the chance of incorrect or misleading entries being injected into the prompt.

Review the glossary section

The generated prompt includes only the document-relevant terms extracted by TermScan from your enabled termbases. Check that the glossary accurately reflects your terminology preferences. You can ask the AI to reorganise terms by category or add missing mappings.

Caution

If your termbase contains incorrect or low-quality entries, these will be injected into the prompt and the AI will be forced to follow them. Only enable termbases that you trust. When starting a new project with no established terminology, consider disabling termbases entirely and letting the AI translate freely – then add terms as you review.

Use with Batch Translate

After saving the generated prompt, select it from the prompt dropdown on the Batch Operations tab. It works with all scopes and providers, just like any other prompt.

AutoPrompt always uses your configured AI provider

Clipboard Mode does not apply to AutoPrompt – ticking the Clipboard Mode checkbox affects only the actual Translate / Proofread passes, not prompt generation. AutoPrompt always sends the meta-prompt request to whichever provider is selected in AI Settings. This enables a useful pattern: keep Clipboard Mode ticked, click AutoPrompt to generate the prompt via your paid API, then run the bulk Translate via clipboard against a free web-tier model. See Combining with AutoPrompt – the hybrid pattern for the full workflow.

Regenerate when the project changes

If your project evolves significantly (new terminology, different document sections, additional termbases), run AutoPrompt again to generate an updated prompt.

---

See Also

• Batch Translate
• Prompts
• Supervertaler Assistant
• AI Settings