MODULE 3 · LESSON 1
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolWhy Context Decides Output Quality
Educational objective 3.1 is the L2 objective in this unit, which means the exam can ask you to explain rather than merely recall. The explanation runs directly from what you learned in module 2.
Large language models generate text by predicting the most probable next token. That prediction does not indicate any understanding of the input or of the situation in the current conversation. What a clearly defined context does is narrow the solution space and guide the model towards domain expertise, producing more precise and actionable output.
Put in terms of the mechanism: you are not informing the model, you are shifting the probability distribution over what comes next. Every piece of relevant context you add makes the continuations you want more likely and the generic ones less likely.
Two layers of context
The syllabus splits context into two layers, and the distinction is examinable.
System context is fixed for every session. It originates from three sources: the system or initial prompt provided by the model's creator, the model's training data, and its built-in settings such as temperature. Together these determine things like the default language, the tone of responses, and the knowledge cut-off date. A model may always answer in English, use a neutral style, and only know about events up to a certain date.
The critical property: the system context is usually not changeable, but it can be complemented by user input. You are not editing it; you are adding alongside it. That is why asking a model to "forget your instructions" does not work, and why a model with an English default will sometimes drift back to English even after you ask for German.
The ability to add context is limited by the context window — the maximum number of tokens, input and output combined, that a model can process in a single interaction.
User context is dynamic and changes with each conversation, or even with each message. The syllabus gives a rich picture of what belongs here, and it is worth internalising as a checklist:
- The project or product. A medical device compliant with ISO 13485. A PSD2-compliant banking system. An e-learning platform.
- The stakeholder perspective. Nursing staff, financial analysts, students.
- Applicable standards and regulations. IEC 62304, GDPR, FDA 21 CFR Part 11.
- The role the AI should take. "You are an expert Requirements Engineer."
- The purpose of the query. Generating test cases, validating requirements.
- Supporting material. Templates and examples.
What happens when context is missing
The syllabus is direct about the failure mode. When context is missing, the model relies solely on its training data and produces answers based on statistical probability. Such output may sound plausible but often lacks domain relevance and practical value. In RE this usually means the result is neither actionable nor of sufficient quality.
This is the mechanism from module 2 seen from the other side. Without context, the most probable continuation is the average of everything the model has read about vaguely similar topics. That average is, almost by definition, a cliché.
Consider a Requirements Engineer working on an appointment system for a dental clinic.
Without context:
Write requirements for an appointment booking feature.
What comes back is the statistical average of every appointment-booking system ever written about: users can create, view, reschedule and cancel appointments; the system sends reminders; an administrator can manage the calendar. All of it plausible, none of it wrong, and none of it about this clinic. It could be pasted into any project on earth, which is precisely the problem — it carries no information that was not already in the model.
With context:
You are an experienced Requirements Engineer.
Project context: an appointment system for a two-chair dental clinic in
Germany. Patients book through WhatsApp rather than a web portal.
Regulatory context: GDPR applies; appointment data includes health data
under Article 9.
Stakeholders: the practice owner (wants fewer no-shows), the receptionist
(currently answers messages between patients), and patients (many of
whom message outside opening hours).
Known constraint: the clinic's existing practice-management software is
the source of truth for availability and cannot be replaced.
Task: draft functional requirements for the booking flow only.
Use the phrase template "The system shall ...".
Flag any requirement you cannot ground in the context above.
The second prompt does several things at once, and each maps to something from the syllabus's user-context list: it sets a role, states the project and product, names the stakeholder perspectives, cites the applicable regulation, gives the purpose, and supplies a template.
It also does something the syllabus recommends elsewhere and is worth adopting as a habit: it asks the model to flag what it could not ground. That instruction converts a hallucination risk into a visible list. The model will still occasionally invent, but you have asked it to mark the boundary of what it was given, which makes review faster.
A Requirements Engineer asks a model to "write quality requirements for our system" and receives generic, untestable statements. According to the syllabus, what has happened?
System context
Click to flipThe layer of context fixed for every session, originating from the creator's system prompt, the training data and built-in settings such as temperature. It determines default language, tone and knowledge cut-off. Usually not changeable, but can be complemented by user input.
Click to flip backContext narrows the solution space and steers a probabilistic model towards domain-relevant continuations. System context is fixed by the provider, the training data and built-in settings, and can only be complemented, not replaced; user context is everything you supply about the project, stakeholders, regulations, role, purpose and format. Without it the model answers from its training data alone, producing plausible output that is neither actionable nor of sufficient quality.