START HERE

What you will learn

  • Write instructions that describe a recognisable task.
  • Separate source material from the instructions about using it.
  • Improve a prompt by checking several outputs rather than guessing.

Before you start: No coding required. You need a task and a short, approved sample of its input.

Why “make this better” is hard to judge

The model cannot reliably infer what “better” means to you. A shorter email, a more cautious recommendation and a more persuasive proposal are different tasks. State who the result is for and what it must help them do.

Think of a prompt as a brief you could hand to a capable assistant who has not attended your meetings. They need the purpose, relevant facts, constraints and required format. A role such as “act as an expert” may influence style, but it does not supply missing evidence or guarantee accuracy.

Instructions and source material do different jobs

Instructions describe what to do. Source material contains information to work from. Keep those two parts visibly separate, using headings such as “Task” and “Source notes”.

Material copied from a document or website may itself contain instructions. For example, a retrieved page could say “ignore earlier instructions”. Treat that as document content, not authority over your application. A well-written prompt helps clarify the boundary, but sensitive tool permissions and data access must also be controlled outside the model.

Step-by-step process

  1. State the purpose

    Describe the specific result and its reader: “Draft a short decision brief for a stakeholder deciding whether to run a prototype.”

  2. Provide the information it may use

    Include the approved notes or passages needed for the task. Say whether outside knowledge is allowed. Remove information that you do not have permission to share.

  3. Specify the output

    Name the sections, approximate length and fields you need. A fixed structure makes omissions easier to notice than an unrestricted essay.

  4. Describe uncertainty and limits

    Tell the model what to do if the notes do not answer a question. Asking it to write “not established in the supplied notes” is more useful than forcing a confident conclusion.

  5. Test and revise one thing at a time

    Run several examples. If the model repeatedly invents dates, add an explicit date rule and retest the same examples. Keep the prompt version and the reason for the change.

FOLLOW ONE EXAMPLE

Worked scenario

Turning research notes into a decision brief

In this fictional example, approved notes describe a search prototype and a manual checklist. The prototype has an estimated build cost but no confirmed delivery date. You want a clear comparison without invented details.

  1. Ask for the same four fields for each option: expected benefit, stated cost, evidence and unanswered questions.
  2. Check whether each claim is supported by the notes. In particular, the missing delivery date should remain an unanswered question.
  3. Try another set of notes with conflicting costs. See whether the output identifies the conflict instead of quietly choosing a number.

Example instruction

Task: Compare the two options for a stakeholder deciding what to investigate next.
Use only the source notes below.
Output: benefit, stated cost, supporting note and unanswered questions for each option.
Do not invent dates or costs. If information is absent, say so.
Source notes:
Option A: Search prototype over approved documents. Estimated build cost: £800. Delivery date: not confirmed.
Option B: Manual evidence checklist. Estimated setup cost: £200. Can start next week.
Ongoing operating costs have not been established for either option.

What this shows: The useful output is a structured first draft with visible gaps. You still make the recommendation after checking the sources.

The process at a glance

The prompt defines the job and available evidence. The source check tells you whether the resulting draft is useful.
ProcessDecisionHuman checkpoint
Read Mermaid diagram source
flowchart LR
 A[Purpose] --> B[Approved source material]
 B --> C[Required output fields]
 C --> D[Model draft]
 D --> E[Compare with the sources]
 E --> F[Revise the instruction]

What you should have at the end

  • A reusable instruction that is specific enough to evaluate.
  • A set of example inputs with notes about where the prompt succeeds or fails.
  • A clearer distinction between what the source says and what still needs investigation.

Common mistakes and how to avoid them

Adding impressive-sounding roles instead of evidence

“World-class strategist” does not tell the model what facts it may rely on.

Changing several instructions after every result

You will struggle to tell which change improved the output.

Treating a good format as a correct answer

A neat table can still contain unsupported claims. Check its contents.

APPLY THE IDEA

Try it yourself

Your task: Take one vague prompt you have used and rewrite it with purpose, source material, output fields and a missing-information rule.

What to produce: Two prompt versions and a comparison of their answers on three different inputs.

Check your understanding: Can you identify a specific improvement or remaining failure, rather than simply saying one answer sounds better?

Key terms in plain English

Prompt
The instructions and context supplied to a model for a task.
Constraint
A rule the answer must follow, such as using only supplied notes.
Prompt injection
An attempt to make untrusted content act as instructions to the system.