It is a quiet Thursday afternoon and Lena, a German-into-English medical translator with eleven years behind her, has a chatbot open in one browser tab and a CAT tool, the computer-assisted translation environment where she edits target text segment by segment, open in the other. The agency just sent her a batch of device labelling for an insulin pump, and a younger PM has been raving about how the new language model "translates better than the engine." So Lena tries it the way most people try it. She pastes a German sentence into the chat box and types four words above it: "Translate this into English." The model answers instantly, in clean, confident, perfectly grammatical English. It reads beautifully. It also silently converts a comma decimal into a period decimal, drops the word "nicht" from a contraindication so the warning now says the opposite of the source, renders the client's approved term "Infusionsset" as the more common "infusion line," and does all of it without a flicker of hesitation, because nothing in her four-word prompt told it that any of those things mattered. This lesson is about those missing instructions. It is about the four levers that separate a generic chatbot guessing at your intent from a disciplined translation assistant that knows it is working against a source, bound to a glossary, fixed to a locale, and required to tell you when it is unsure. Lena's four words got her a fluent stranger. The same model, given the right four levers, becomes something she can actually work with.

Why "Translate This" Is the Worst Prompt You Can Write

Before we build the good prompt, we have to understand exactly why the bare one fails, because the failure is not random and it is not the model being "bad at German." The failure is structural, and it comes from what a prompt actually is. A prompt is the entire instruction set you hand the model: the words you type, plus anything you attach, plus whatever the system has been told to do before your message ever arrives. The model has no other source of intent. It cannot see the client brief sitting in your inbox, it does not know the content is regulated medical labelling, it has never met the termbase, and it has no idea which English-speaking market the file is for. Everything it knows about your job is in the prompt. When the prompt is "Translate this into English," the model is not given a job. It is given a vacuum, and it fills the vacuum with the single most statistically probable thing: smooth, generic, middle-of-the-road English that reads well to the largest number of readers.

This is the trap that the whole first level of this program drilled into you, now showing up in a new place. An LLM (large language model, a general text-prediction system that translates as a side effect of its broad training) is optimized for fluency first and faithfulness to your specific source second. A bare prompt removes every constraint that would force it toward faithfulness and leaves only the pressure toward fluency. The result is output that is maximally pleasant to read and minimally accountable to the source segment in front of it. The cleaner it reads, the more dangerous it is, because fluency is exactly the quality your trained eye reads as a sign that a sentence is fine.

A prompt is not a request for a translation. It is the complete job specification, and everything you leave out, the model fills in with the most generic guess available.

Consider what Lena's four words did and did not communicate. They communicated the target language. That is all. They did not say this is for a US market, so use US conventions. They did not say a comma in the German number is a decimal separator, so 2,5 mg is 2.5 mg and not 25. They did not say "Infusionsset" has an approved English equivalent that the client audits for. They did not say this is a contraindication where a dropped negation can harm a patient, so preserve every negation exactly and flag anything ambiguous. The model could not have known any of that, and so it did the only thing it could: it guessed, fluently and confidently, and three of its guesses were wrong in ways that read perfectly.

The Four Levers, Named

The fix is not a magic phrase or a clever trick. It is the deliberate supply of the four pieces of context the model was missing, each one a lever that pulls the output away from generic fluency and toward disciplined, source-faithful translation. Throughout this lesson we will build them one at a time and watch each one change the output:

• Context. What is this content, who reads it, what register and purpose does it serve, and what are the consequences if it is wrong. This tells the model what kind of correctness it is being held to.
• Glossary and termbase. The client's approved terms, paired source-to-target, that must be used exactly and never paraphrased into a synonym the model "prefers." This is the difference between consistent terminology and term drift.
• Locale. The specific target market and its conventions: number and date formats, units, currency, formality, spelling. A locale is not just a language; it is a language plus a place plus a set of rules.
• Stay faithful to the source. The explicit instruction to translate what is written and not to improve, summarize, infer, or invent, paired with the instruction to flag uncertainty rather than paper over it with confident prose.

None of these is exotic. Every one of them is something you already know as a professional and would naturally apply if you were translating the segment yourself. The entire skill of prompting for a linguist is the discipline of writing down what you already know, because the model cannot read your professional mind and will substitute a generic guess for every piece of judgment you fail to make explicit.

Lever One: Context, the Frame the Model Translates Inside

Context is the lever that tells the model what game it is playing. The same German sentence is a different translation problem depending on whether it is a marketing tagline, a legal clause, a user-interface string, or a line on a drug label, and a bare prompt gives the model no way to know which. Supplying context is the act of handing the model the frame a human translator carries automatically: the document type, the audience, the register, the purpose, and the stakes.

Watch what it does to Lena's job. Here is the weak prompt and the model's reply.

Before: No Frame

Prompt: "Translate this into English: Das Infusionsset nicht länger als 72 Stunden verwenden."

Model output: "Use the infusion line for no longer than 72 hours."

This reads fine. It is grammatical, natural, and confident. It is also wrong in two ways Lena cannot see from the English alone. "Infusion line" is not the client's approved term, and because no context told the model this was clinical labelling held to a fixed terminology, it chose the more colloquial phrase. More subtly, the register is casual for a medical device instruction that, in a regulated US labelling context, would conventionally read as a directive ("Do not use the infusion set for longer than 72 hours"). The model defaulted to friendly consumer English because nothing told it the content was regulated.

After: The Frame Supplied

Prompt: "You are translating regulated medical device labelling for an insulin infusion pump, from German into English. The audience is a patient self-administering insulin at home. The register is formal, directive, and clinical, matching the conventions of US FDA-regulated device instructions. Accuracy is safety-critical: a mistranslation can cause a dosing error. Translate the following sentence and preserve its meaning exactly. Das Infusionsset nicht länger als 72 Stunden verwenden."

Model output: "Do not use the infusion set for longer than 72 hours."

The single change here is that the model now knows what it is translating. The register tightened to a directive, the negation is handled with the care the framing demanded, and the model is primed to treat the term carefully even before we hand it a glossary. Context did not fix everything; it set the stage so the other three levers could land. The word "regulated," the phrase "safety-critical," and the audience description each pull the model's probability distribution toward the careful, conventional, faithful rendering and away from the breezy generic one.

There is a deeper principle here worth stating plainly. The model is always going to produce the most probable output given the prompt. You cannot turn that off. What you can do is change what "most probable" means by changing the prompt, because the probability is conditioned on everything you supply. Saying "this is regulated medical labelling for a patient at home" makes careful, formal, faithful English more probable than breezy consumer English, simply by telling the model which neighbourhood of its training to draw from. Context is not decoration. It is the steering.

How Much Context Is Enough

A reasonable worry at this point is that you will spend more time writing the prompt than translating the segment yourself. For a single sentence, that may be true, and a single sentence is rarely where you would reach for an LLM. The economics work when the context is written once and reused across a batch, a file, or a whole project, which is exactly how a system prompt operates and what a later lesson in this chapter builds toward. For now, the rule of thumb is this: supply the context that changes the translation. Document type, audience, register, and consequence almost always change it. The client's office address almost never does. You are not writing a brief for a new hire; you are supplying the specific facts that move the output, and you learn which facts those are by watching what the model gets wrong without them.

Lever Two: Glossary and Termbase, the End of "Close Enough"

The second lever attacks the failure mode that quietly costs agencies the most rework: term drift. An LLM does not have a concept of "the approved term." It has a concept of "the most probable word for this meaning," and the most probable word is the common one, not necessarily the one your client has spent years standardizing in their documentation, their UI, their training materials, and their regulatory filings. Left alone, the model will render the same source term three different ways across a file, every rendering fluent, every one a terminology error that an evaluator scoring against an MQM (Multidimensional Quality Metrics) error typology will flag, and that a client running an automated terminology check will reject.

A termbase (the database of approved source-to-target term pairs, sometimes called a glossary in looser usage) is the cure, and the way you deliver it to the model is by putting the relevant pairs directly into the prompt as hard constraints. You are not asking the model to consider the glossary. You are telling it that these mappings are non-negotiable and that any deviation is an error.

Before: The Model Picks Its Favourite Word

Prompt: "Translate these three sentences into English, US market, regulated medical labelling: Das Infusionsset anschließen. Das Infusionsset alle 72 Stunden wechseln. Bei Verstopfung das Infusionsset entfernen."

Model output:

• "Connect the infusion line."
• "Replace the infusion kit every 72 hours."
• "In case of a blockage, remove the infusion device."

Read those three English sentences and the problem leaps out: "infusion line," "infusion kit," "infusion device," three renderings of the single German term Infusionsset across three consecutive sentences. Each one is fluent. Each one is defensible in isolation. Together they are a terminology disaster, because a regulated client requires one approved term used consistently, and inconsistency in a medical device label is not a style quibble; it can read as referring to three different components. The model was not being careless by its own lights. It was doing exactly what it optimizes for, varying its word choice for natural-sounding prose, which is the right instinct for an essay and the wrong instinct for terminology-controlled content.

After: The Termbase as a Hard Constraint

Prompt: "Translate these three sentences into English for the US market. This is regulated medical device labelling and accuracy is safety-critical. You MUST use the following approved terminology exactly, with no synonyms or variations:
German 'Infusionsset' = English 'infusion set' (always; never 'line', 'kit', 'tube', or 'device').
German 'Verstopfung' = English 'occlusion' (always; never 'blockage' or 'clog').
If a sentence uses an approved source term, the corresponding approved target term must appear. Sentences: Das Infusionsset anschließen. Das Infusionsset alle 72 Stunden wechseln. Bei Verstopfung das Infusionsset entfernen."

Model output:

• "Connect the infusion set."
• "Replace the infusion set every 72 hours."
• "In the event of an occlusion, remove the infusion set."

The term now holds across all three sentences, and "Verstopfung" resolves to the clinically correct "occlusion" rather than the lay "blockage." Three things made this work, and each is a discipline you should internalize. First, the pairs were given explicitly, source and target, not just a list of English words the model had to guess where to apply. Second, the forbidden alternatives were named ("never 'line', 'kit', 'tube', or 'device'"), which is far more effective than a positive instruction alone because it closes the specific escape routes the model would otherwise take. Third, the instruction was framed as a hard rule ("you MUST," "always") rather than a polite suggestion, because the model weights a stated constraint more heavily when it is unambiguous about being mandatory.

The model does not know your client's approved term exists. If you do not put the termbase in the prompt as a hard constraint, the model will use the most common word, fluently, every time, and call it a translation.

The Limit of the Glossary in the Prompt

Honesty about the limits matters, because over-trusting this lever is its own failure mode. Putting terms in the prompt works well for a handful of terms on a short batch. It degrades as the term list grows into the hundreds and the source text grows long, because the model can lose track of a constraint buried far up in a large prompt, and it has no mechanical lookup the way a CAT tool's terminology checker does. The prompt-level glossary is a strong first control, not a guarantee. You still verify the output against the termbase, because the model can acknowledge a rule and then violate it three sentences later with complete confidence. Later lessons in this program build the grounded, retrieval-based and tool-enforced approaches that scale past what a prompt alone can hold. For now, the lesson is that even the simple act of pasting the relevant term pairs into the prompt eliminates the most common and most embarrassing class of terminology error, and costs you thirty seconds.

Lever Three: Locale, Because a Language Is Not a Market

The third lever corrects an error so common it is almost invisible: treating a language as if it were a place. "English" is not a translation target. en-US and en-GB are translation targets, and they differ in spelling, in date and number formats, in units, in currency, in vocabulary, and in register, in ways that matter intensely on exactly the high-consequence content where the cost of error is highest. A locale is the technical term for a language paired with a region and its conventions, written as a code like de-DE, en-US, pt-BR, or fr-CA. When you tell the model "English," you have given it a language and left the locale blank, and it will fill that blank with whatever its training found most probable, which is usually a vague mid-Atlantic English that belongs to no actual market.

For Lena's insulin pump, the locale stakes are not cosmetic. They are clinical. Watch the difference.

Before: "English," Locale Unspecified

Prompt: "Translate into English: Verabreichen Sie 2,5 mg am 03.06.2026. Die Patientin wiegt 70 kg."

Model output: "Administer 2.5 mg on 03.06.2026. The patient weighs 70 kg."

The dose converted correctly here, but look at the date: "03.06.2026" was carried straight through unchanged. In the German source, written day-first, that is the 3rd of June. A US reader of "03.06.2026" will read it month-first as the 6th of March, three months earlier. The model preserved the digits and corrupted the meaning, because no locale told it to convert the date format. On a medication schedule, a date that reads as a different day is precisely the silent critical error this program exists to catch: fluent, confident, and clinically dangerous.

After: en-US Specified, Conventions Spelled Out

Prompt: "Translate from de-DE into en-US. Apply US locale conventions: convert dates from German day-month-year format to US month/day/year format and write the month name in full to avoid ambiguity; use a period as the decimal separator; keep metric units (mg, kg) as they appear since this is clinical dosing, but ensure number formats follow US convention. This is medical content where a date or number error is safety-critical. Source: Verabreichen Sie 2,5 mg am 03.06.2026. Die Patientin wiegt 70 kg."

Model output: "Administer 2.5 mg on June 3, 2026. The patient weighs 70 kg."

The date is now unambiguous and correct for the US reader, the decimal is right, and the units were deliberately kept because the prompt told the model that clinical dosing keeps metric. Notice the subtlety in that last instruction: locale conversion is not a blanket "make everything American." A naive locale rule might have helpfully converted 70 kg to 154 lbs, which on a clinical dosing label would be a catastrophic, unrequested transformation. The good prompt specified exactly which conventions to apply and which to leave alone, because locale correctness is a set of specific rules, not a vibe, and the consequences run in both directions: failing to convert a date is an error, and over-converting a clinical unit is a worse one.

A language is what the model speaks. A locale is the market it speaks to, with its own rules for dates, numbers, units, and formality. Specify the locale, and specify which conventions apply, because "English" alone is a guess waiting to go wrong.

Locale Carries Formality, Too

Locale is not only about formats. It carries register and formality conventions that have no equivalent in the source language's grammar. German distinguishes the formal Sie from the informal du; English has no grammatical equivalent, but the choice still has to be rendered somehow, through word choice, contraction, and tone. A locale instruction can carry that: "Use a formal, respectful register appropriate for patient-facing medical instructions in the US market; avoid contractions." Languages like Japanese, Korean, and many others encode honorifics and formality levels far more elaborately, and the target locale's conventions decide how those map. A model with no locale instruction will pick a default formality, usually a friendly, contraction-heavy, mid-formal English, that may be exactly wrong for a legal contract or a clinical label. The locale lever, fully used, sets format and register together.

Lever Four: Stay Faithful to the Source, and Flag What You Cannot Be Sure Of

The fourth lever is the one that most directly addresses the discipline's deepest fear, and it has two halves that work together. The first half is an explicit instruction to stay faithful to the source: translate what is written, do not improve it, do not summarize it, do not infer missing intent, do not add helpful clauses, do not invent. The second half is the instruction that turns the model from a confident liar into a useful assistant: when you are uncertain, say so, rather than producing fluent prose that hides the uncertainty.

The first half matters because an LLM, far more than a narrow translation engine, is a willing improver. Its training rewards helpfulness, and "helpfulness" applied to a confusing source segment looks like guessing the author's intent and rendering that guess in authoritative prose. On creative content that instinct can be an asset. On a contract, a contraindication, or a financial disclosure, it is a liability, because the model will smooth an ambiguous source into a confident, specific target that asserts more than the source actually said. The faithfulness instruction is how you tell the model that its job here is to convey, not to author.

Before: The Helpful Improver

Prompt: "Translate into English: Bei Bedarf kann die Dosis angepasst werden."

Model output: "If needed, the dose can be increased."

The German says the dose "can be adjusted" (angepasst), which means changed in either direction, up or down. The model rendered it as "increased," narrowing an open instruction into a specific and unsupported one. Why? Because in a great deal of medical training text, dose adjustments trend toward increases, so "increased" was the more probable completion, and the model, helpfully, asserted it. On a label, this is the difference between a safe instruction and a directive that could lead a patient to raise a dose the source never told them to raise. The model did not hesitate. It produced a fluent, specific, wrong sentence, because nothing told it to stay literal and nothing told it to flag the ambiguity.

After: Faithful, and Honest About Uncertainty

Prompt: "Translate from de-DE into en-US. Translate the source exactly as written. Do not add, remove, narrow, or infer meaning that is not explicitly present in the source. If the source is ambiguous, or if a term could reasonably translate more than one way and the choice affects meaning, do not silently pick one: translate it as faithfully as possible and add a bracketed note flagging the ambiguity and your reasoning, so a human reviewer can resolve it. Source: Bei Bedarf kann die Dosis angepasst werden."

Model output: "If needed, the dose can be adjusted. [Translator note: 'angepasst' means 'adjusted' and is directionally neutral; it could be changed up or down. I have kept the neutral 'adjusted' rather than narrowing to 'increased' or 'decreased'. Confirm intended direction with the source author if a specific direction is required.]"

This is the single most valuable behaviour you can elicit from the model, and it is worth pausing on why. The model did not become more accurate at translation; it became honest about the limits of what it could determine. The neutral rendering is now correct, and the bracketed note hands Lena exactly the thing a confident output would have buried: a flag that says "here is a decision a human needs to make, and here is why." A model that flags uncertainty has converted a silent critical error into a visible query, and a visible query is something a post-editor can resolve in seconds. The whole purpose of human-in-the-loop quality is to catch the dangerous segment before it ships, and a model instructed to surface its own doubt is doing your most important detection work for you.

An unprompted model hides its uncertainty inside fluent prose. A model told to flag uncertainty turns the silent critical error into a visible query, which is the single most useful thing it can do for you.

Why a Flag Beats a Confident Guess Every Time

There is a temptation to see the bracketed note as the model "failing" to produce a clean translation, and to prefer the version that just answered. Resist it completely. The confident version that said "increased" looked cleaner and was wrong in a way that could harm a patient and would never have been caught by reading the English. The flagged version looked messier and was right, and it routed the one genuinely ambiguous decision to the one party qualified to make it. In high-consequence localization, a visible question is always worth more than an invisible assertion. The model that says "I am not sure, here is why, please check" is behaving the way a junior translator should behave: surfacing the hard call instead of hiding it. You are not training the model to be uncertain about everything, which would make it useless. You are training it to distinguish what it can render confidently from what genuinely turns on a human decision, and to tell you which is which.

This faithfulness-plus-flagging lever also interacts with the other three. Context tells the model the stakes are high, which makes it appropriately more cautious about where it flags. The glossary removes terminology ambiguity, so the flags that remain are real meaning ambiguities rather than word-choice noise. The locale removes format ambiguity. With the first three levers in place, the fourth lever's flags become a clean, high-signal list of exactly the segments a human must adjudicate, which is precisely the routing that makes an AI-assisted workflow both fast and safe.

Putting the Four Levers Together

Each lever was shown alone so you could see its specific effect. In real work you stack all four into one prompt, and the combination is what turns the chatbot into the disciplined assistant the title promised. Here is Lena's prompt with everything assembled, on the same source sentence that opened the lesson, the one her four-word prompt mangled three different ways.

The Assembled Prompt

Prompt: "You are a translation assistant for regulated medical device labelling. Translate from de-DE into en-US.
CONTEXT: This is patient-facing labelling for an insulin infusion pump used by patients at home. Register is formal, directive, and clinical, matching US FDA-regulated device-instruction conventions. Accuracy is safety-critical; a mistranslation can cause a dosing error.
TERMINOLOGY (mandatory, exact, no synonyms): 'Infusionsset' = 'infusion set' (never line/kit/tube/device); 'Verstopfung' = 'occlusion' (never blockage/clog).
LOCALE: Convert dates to US month/day/year with the month spelled out; period as decimal separator; keep metric clinical units (mg, kg) unchanged; formal register, no contractions.
FIDELITY: Translate exactly as written. Do not add, remove, narrow, or infer meaning. If anything is ambiguous or could change meaning, translate faithfully and add a bracketed [Translator note] flagging it for human review rather than guessing silently.
Source: Das Infusionsset nicht länger als 72 Stunden verwenden."

Model output: "Do not use the infusion set for longer than 72 hours."

The negation is preserved, the approved term is correct, the register is right for a regulated label, and there is no spurious flag because nothing in this particular sentence was ambiguous. The same model that, given four words, dropped a negation and drifted the term and corrupted a number, now produces a clean, on-term, locale-correct, faithful rendering, because it was finally told what job it was doing. Nothing about the model changed between Lena's first attempt and this one. Everything about the instruction changed.

The Prompt Is Not the Quality Gate

One caution must close this section, because it is the line that separates a professional from someone who has learned a trick. A good prompt makes good output more probable. It does not make good output certain. The model can still, despite a perfect prompt, drop a negation, drift a term, or carry a wrong number, because it is a probability machine and not a rule engine, and "more probable" is not "guaranteed." The four levers raise the floor dramatically; they do not remove the requirement to verify. Lena still reads the output against the source segment, still checks the negation, the number, the date, and the term, still owns the delivery. The prompt is how you give yourself the best possible draft to verify. The verification, and the accountability for the shipped string, remain exactly where the entire program insists they belong: with the human whose name is on the file. The prompt is a better starting line, not a finish line.

This is also why the very next lessons in this chapter take you deeper into each lever: forcing the model to the source segment instead of its own fluency, recognizing the specific patterns of bad output, and eventually locking these instructions into a reusable system prompt so you do not retype them on every file. The four levers are the foundation. The discipline of verifying what they produce is the building you put on top of it.

Key Takeaways

• A prompt is the complete job specification you hand the model, and the model has no other source of intent: it cannot see the client brief, the termbase, the regulated context, or the target market unless you put them in the prompt, so everything you leave out it fills with the most generic, most fluent guess available.
• "Translate this into X" fails structurally because an LLM (large language model) is optimized for fluency first and faithfulness to your specific source second; a bare prompt removes every constraint that would pull it toward faithfulness and leaves only the pressure toward smooth, generic, unaccountable prose.
• Lever one, context, frames the translation: document type, audience, register, purpose, and consequence change the output by changing what "most probable" means, steering the model toward the careful, conventional rendering and away from the breezy default.
• Lever two, the glossary and termbase, defeats term drift by putting approved source-to-target pairs into the prompt as hard constraints, naming the forbidden alternatives explicitly ("never line/kit/tube"), and framing them as mandatory; it is a strong first control, not a guarantee, and still requires verification against the termbase.
• Lever three, locale, corrects the error of treating a language as a place: en-US and en-GB differ in dates, numbers, units, currency, spelling, and formality, so you specify the locale code and the exact conventions to apply, and just as importantly the ones to leave alone, since over-converting a clinical unit is worse than not converting it.
• Lever four, stay faithful to the source plus flag uncertainty, stops the model from helpfully improving, narrowing, or inventing, and instructs it to surface ambiguity as a visible bracketed query rather than burying it in confident prose; a flagged uncertainty is a silent critical error converted into a question a human can resolve in seconds.
• The four levers stack into one prompt that turns a generic chatbot into a disciplined translation assistant, and they reinforce each other: context sets the stakes, the glossary removes term ambiguity, the locale removes format ambiguity, and the fidelity lever's flags become a clean, high-signal list of exactly the segments a human must adjudicate.
• A good prompt raises the floor; it is not the quality gate. The model remains a probability machine that can still drop a negation or drift a term despite a perfect prompt, so the human reads the output against the source segment, verifies numbers, negations, dates, and terms, and owns the delivery, exactly where accountability belongs.