PMGM7023 Chap.5 Prompts, Context and the Model Engine
Prompts, Context and the Model Engine
A vocabulary arranged so failures can be located
The second session supplies eighteen terms arranged in five layers, and the arrangement carries as much teaching as the definitions. Each layer answers one question about the system, so a failure can be attributed to a layer rather than blamed on the model in general.
This chapter covers the first two layers: what enters the system, and what generates the response.
A request can specify how to think, not only what to answer
A prompt is your request together with anything you deliberately supply alongside it. A strong one names the outcome wanted, the context, the evidence available, the constraints, the reasoning approach and the shape of the output.
Each of those removes a degree of freedom that would otherwise be filled by guessing.
Asking what should we do supplies one element; asking what should we do, given these two tables, without inferring causation, as three findings and one recommended next question supplies five.
Preferences that apply across many tasks belong in a standing instruction instead: who you are and what you work on, how you want answers organised, and the habits you never want, including evidence rules and hard constraints.
The test of a good one is whether it would visibly change an answer.
A reusable request works the same way, holding fixed the standards of reasoning, the rules on evidence, the shape of the output and the final check while leaving obvious places for the goal, the audience and the source material.
Context is finite and it competes
Context is everything available for the current response: the standing instruction, the conversation so far, attached files, passages found for you, and results returned by tools.
Two limits follow. The first is capacity, and anything falling outside it cannot be used at all rather than merely being deprioritised, which is the mechanism behind an assistant that appears to forget an instruction given an hour earlier.
The second is attention: a long input can bury the passage that decides the answer, so supplying a hundred pages when four are relevant makes a wrong answer more likely and harder to diagnose.
The sampling setting is frequently misread as an accuracy control. A lower value concentrates probability on the most likely continuations and usually produces more repeatable wording.
It adds no knowledge, verifies no claim and does not reduce the chance of an invented fact.
What is generating the response
Generative systems are the broad category producing new content, including images and audio, so the category is wider than language.
A large language model is the language member of that family and can power a product without being the product, since the product may also include search, tools, memory and controls.
The architecture asks, for each position in a sequence, which earlier positions carry useful signal, merges what it finds and revises how it is holding that position, repeating that block across many layers before scoring and choosing the next piece of the answer.
Two moments are easy to conflate. Training adjusts numerical parameters before deployment.
Inference uses an already-trained model on new input, which is what a tutorial request triggers, so a correction you make does not change the model and will not persist tomorrow unless something outside it stores the correction.
What this chapter covers
- 01
Prompt, standing instruction and reusable request as three different objects
- 02
The context window as a finite space with competing occupants
- 03
Sampling spread changes variation in wording, not accuracy
- 04
Training changes parameters; inference uses them
Rebuilding a request so the answer can be judged
- 3Name the decision the summary will support.
- 3Name the audience and the complication the answer must handle.
- 3State an evidence rule and a refusal in advance.
Key terms
- Standing Instruction
- A standing instruction is a persistent set of preferences and rules read at the start of a session and applied across tasks. It is useful only if it would visibly change an answer.
- Context
- Context is all information available to a model for the current response, including instructions, conversation history, attached files and tool results.
- Inference
- Inference is the use of an already-trained model on a new input to produce an answer. It leaves the model's parameters unchanged.
- Temperature
- Temperature is a generation setting that reshapes how probability is spread across candidate continuations. It changes variation in wording rather than correctness.
Prompts, Context and the Model Engine FAQ
Where should I put a preference that applies to every task?
In the standing instruction rather than in each request. A preference repeated message by message consumes space in a finite working context and disappears as soon as earlier turns fall outside it. Written once as a standing instruction, it is read at the start of a session and applies across tasks, which is both cheaper and more reliable than restating it.
Will lowering the temperature stop an assistant inventing figures?
No. The setting reshapes how probability is spread across possible continuations, so it affects how varied and repeatable the wording is. An invented figure is a missing-evidence problem: the repair is to supply the passage containing the number, require every figure to be quoted from the attachment, and permit the answer to state that the figure is not present.
Exam move
Write your standing instruction early in the term and revise it whenever you correct the same habit twice. Keep tutorial sessions short and start fresh when the task changes. Practise naming the layer responsible for a described failure before proposing any fix, since most failures are input or evidence problems rather than model problems.
Working through Prompts, Context and the Model Engine in PMGM7023? Sia is AskSia’s AI Management tutor — ask any PMGM7023 Prompts, Context and the Model Engine question and get a clear, step-by-step explanation grounded in how PMGM7023 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
