Skip to content

Prompting and Instructions — Junior

At junior level, focus on this question:

Can you write a prompt with the five parts every effective prompt has — and explain why vague instructions produce vague output?

The five parts

  1. Role: who the model is — "You are a support agent for an e-commerce company." Focuses vocabulary and tone.
  2. Task: the one thing to do, stated as an action — "Classify this ticket into exactly one category." One prompt, one primary task.
  3. Context: what the model needs to know for this request — the ticket text, the customer's plan tier. Only what's needed; context costs tokens on every call (see Tokens and Context).
  4. Constraints: the boundaries — what to do, what never to do, how to handle missing information ("If the order ID is missing, ask for it; don't guess").
  5. Output format: exactly what the response should look like — "Reply with JSON: {category, confidence}" — or "Reply in 2 sentences maximum."
flowchart TB P["Prompt"] --> R["Role: who you are"] P --> T["Task: do this one thing"] P --> C["Context: here's what you need"] P --> K["Constraints: never do this / handle gaps like this"] P --> F["Format: respond exactly like this"]

Why vague prompts fail — the probability view

  • The model predicts the most plausible continuation of what you wrote. A vague prompt is compatible with many continuations; the model picks one, and it may not be yours.
  • "Summarize this" → plausible for a paragraph, a bullet list, a tweet-length gist — the model guesses your intent. "Summarize in 3 bullet points, each under 15 words, for a busy manager" → one dominant continuation.
  • Specificity isn't pedantry — it's how you collapse the distribution onto the output you actually want.

Show, don't tell

  • "Be concise" is weaker than "Reply in 2 sentences maximum."
  • "Don't be verbose" is weaker than a one-line example of the desired length and shape.
  • Models copy patterns better than they follow abstractions — one concrete example outweighs three adjectives.

Common Mistakes

  • One vague instruction doing five jobs. "Write a good email" — good how, to whom, about what, how long?
  • Burying the task mid-paragraph. Put task and format where they're unmissable (first or last, not the middle of a wall of context).
  • Constraints without gap-handling. Telling the model what to do on happy paths only; unstated edge behavior gets improvised.
  • Politely hoping. "Please try to maybe..." — instructions are specifications; hedged wording yields hedged compliance.

Apply It

  1. Take one prompt you use; label every sentence with its part (role/task/context/constraint/format) — and write the missing parts.
  2. Rewrite its vaguest instruction as a concrete, checkable rule ("be concise" → "2 sentences max").
  3. Run both versions on 3 real inputs and compare outputs against your intent.

Verify Your Work

  • The prompt has all five parts, each doing one job.
  • Every abstract instruction ("good", "concise", "professional") is replaced by something checkable.
  • The format section describes the exact response shape, including the edge cases.

Review Questions

  • Why does a vague prompt produce unpredictable output, in probability terms?
  • Why is "2 sentences max" stronger than "be concise"?
  • Which two prompt parts do people most often omit, and what does each omission cost?