Server AtticLearning Harbour

AI Foundations · Module 3 · Lesson 1 of 3

Instructions as task design: goal, context and constraints

Writing an AI instruction is a small act of task design, not a request for a favour — this lesson breaks it into three checkable building blocks (goal, context, constraints) and shows what a vague instruction forces the AI to invent instead.

Lesson · 15–20 minutes · Text-first

By the end, you can

  • Identify the three building blocks of a checkable AI instruction — goal, context and constraints (AF-3).
  • Explain why an instruction missing context or constraints forces the AI to fill the gap with an unstated assumption (AF-3, building on Module 1's fluency point and Module 2's judgement-risk point).
  • Rewrite a vague instruction into one that states its goal, supplies context and states its constraints (AF-3).

Before you start

This is Module 3, Lesson 1 of 3, the first lesson in this course's work on writing AI instructions. It builds on Module 1's vocabulary — pattern, model, hallucination — and on Module 2's task-type distinctions and its three-question decision habit for whether a task is safe for AI at all. This lesson assumes the task has already passed that habit: the question here is not whether to use AI, but how to instruct it so the result is worth checking. This lesson does not require coding or technical setup.

A vague instruction is a design decision left unmade

"Write something for the newsletter" is a request. It is not a task design. Every AI instruction quietly hands over a set of decisions: what counts as done, what facts to use, what to leave out. When an instruction states those decisions, the AI can act on them. When it does not, the AI still has to produce something — so it fills the gap with a plausible guess. Anthropic's current prompting guidance makes this concrete with its own before/after: the instruction "Create an analytics dashboard" is described as less effective than "Create an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation" — the second version states the goal and scope the first left for the model to invent.

The same guidance offers a useful test for this: **"Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too."** A vague instruction does not become clearer because the reader is a model instead of a person.

Three building blocks: goal, context, constraints

Three parts turn a vague request into a checkable instruction:

None of the three is a nicety. Leave out the goal and the AI optimises for something unstated. Leave out context and it fills gaps with a plausible-sounding but unverified guess — the same fluency risk Module 1 described in general terms, now showing up because the instruction, not just the AI, left a gap. Leave out constraints and nothing stops it producing an invented fact, the wrong tone, or content outside its bounds.

  • **Goal** — the actual outcome the work is for, stated specifically enough that a reader can tell whether the result achieved it. "Write something for the newsletter" has no goal; "draft a 150-word newsletter item announcing the office move, for staff who have not yet heard" does.
  • **Context** — the material the AI is actually allowed to use: source text, data, prior decisions, the reason the task matters. Anthropic's guidance calls this "context or motivation behind your instructions," and states plainly that "providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals and deliver more targeted responses." Google's prompt design guidance makes the same point from a different angle: "you can include instructions and information in a prompt that the model needs to solve a problem, instead of assuming that the model has all of the required information." OpenAI's prompt engineering guidance states it as a plain instruction-writing habit: "include additional context information the model can use to generate a response within the prompt you give the model."
  • **Constraints** — what must not happen: facts not to invent, a tone to keep, a length limit, a policy or privacy boundary. This is where Module 1's hallucination point and Module 2's judgement-risk point turn into something you can write down in advance, rather than catch after the fact.

A worked example

A facilities coordinator wants an email to staff about an upcoming office move. Their first instruction is: "Write an email about the office move."

Applying the three building blocks: **Goal** — a short, reassuring email confirming the move date and the one action staff need to take (updating their desk-booking profile), for staff who already know a move is happening but not the details. **Context** — the actual confirmed facts: the move is on 14 September, the new floor is 3, and the desk-booking system link is `intranet/desks`. Nothing beyond these three facts is confirmed yet. **Constraints** — do not state a parking or storage policy, since neither has been decided; keep it under 120 words; keep the tone plain and practical, not celebratory.

The rewritten instruction: "Draft a staff email confirming the office move. Goal: staff know the move date and complete one action (updating their desk-booking profile) — this is a reassurance and instruction email, not an announcement of a change they don't already know about. Context: the move date is 14 September, the new floor is 3, and the desk-booking link is intranet/desks — use only these three facts. Constraints: do not mention parking or storage, since neither is decided; keep it under 120 words; keep the tone plain and practical." A person reading this instruction — with no other context — could follow it. That is the same test the AI is being held to.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The practice exercise's model answer sits behind a native disclosure control that is reachable and operable by keyboard and correctly announced by screen readers. The knowledge check uses native radio-button inputs with a visible question and options, and posts its result to a live status region so assistive technology announces the outcome without a page reload.

Practice

Rewrite a vague instruction: the quiz request

A teacher asks an AI tool to 'make a quiz about the chapter.' The chapter is chapter 6 of the class's assigned textbook, on the water cycle. The quiz needs to be ready for a 20-minute activity at the start of tomorrow's class, cover only material the class has actually covered so far (evaporation and condensation, not precipitation or collection, which are next week's topics), and must not include any question the teacher would need to look up the answer to before using it.

  1. Write a specific goal for this instruction — what counts as a usable result, for whom, by when.
  2. Write the context: what source material and scope limits should the AI be told about, rather than left to guess?
  3. Write at least two constraints that follow directly from the scenario.
  4. Rewrite the full instruction in one paragraph, applying all three building blocks.
Compare with a bounded first version

Goal: a ready-to-use 20-minute warm-up quiz for tomorrow's class, testing only what the class has already covered. Context: the quiz covers chapter 6 (the water cycle) but only the evaporation and condensation sections already taught — precipitation and collection are next week's topics and are out of scope; the teacher needs the quiz ready before tomorrow's class starts. Constraints: do not include any question about precipitation or collection; every question must have an answer the teacher can verify directly from the covered material, not one requiring them to look anything up; the quiz must fit a 20-minute activity, so a short format (for example, 8-10 short-answer or multiple-choice questions) fits better than a long one. Rewritten instruction: 'Write an 8-10 question warm-up quiz for tomorrow's 20-minute class activity, covering only the evaporation and condensation sections of chapter 6 (the water cycle) — do not include precipitation or collection, which the class has not covered yet. Every question must be answerable directly from material already taught, with no fact I would need to look up before using it.'

Knowledge check

Try the idea

An instruction says 'Summarize the customer complaint and suggest next steps' but does not say which complaint, does not supply the complaint text, and does not say who the summary is for. What is the most accurate description of what happens next?
Low-stakes practice only. This does not score, block progress or create a learner record.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

  1. Prompting best practices — Anthropic. Recommends being specific about desired output and constraints, and states that providing context or motivation behind an instruction helps the model deliver more targeted responses.
  2. Prompt engineering — OpenAI. States that a prompt should include additional context information the model can use to generate a response.
  3. Prompt design strategies — Google AI for Developers. Recommends including the instructions and information a model needs to solve a problem rather than assuming the model already has it.
  4. Artificial Intelligence Risk Management Framework 1.0 — NIST AI Resource Center. Frames an AI system as an engineered or machine-based system that generates outputs such as predictions, recommendations or decisions — not a system that already knows a task's unstated goal, source material or limits.