A prompt is a task brief for a model. Context, a specific task, the output format, constraints, and quality criteria produce a more predictable result than a supposedly special phrase.

The model does not know hidden context

This is the second article in the series. The first covers model selection and context preparation.

A request such as “do some marketing” tells the model nothing about the product, audience, or available budget. Generic copy is the likely result.

A person given the task “make a presentation” can ask follow-up questions. A model often starts answering immediately, so the prompt should remove the main sources of uncertainty.

Five parts of a working prompt

Context. Describe the company, product, audience, current situation, and source data. A long, structured input is more useful than a short request with missing information.

Task. State the action you expect. Instead of “analyze churn,” ask for three likely causes of fourth-quarter customer churn, a separation of facts and assumptions, and actions within a $3,000 budget.

Output format. Specify whether you need a list, email, table, two-page plan, or presentation outline. Otherwise, the model will choose the form itself.

Constraints. List the boundaries: avoid invented figures, exclude unverified sources, preserve legal wording, or stay within 1,500 characters.

Quality criteria. Explain how you will accept the answer. For example, it includes sources, risks, and alternatives, while keeping conclusions separate from facts.

Microsoft begins its basic guidance with a concise rule:

“Start with clear instructions.” — Microsoft Learn

Source: Microsoft Learn, Prompt engineering techniques

The model no longer has to guess what success means. Clear criteria reduce generic prose and unnecessary iterations.

A role does not replace the task

The phrase “you are a senior marketer with 15 years of experience” can set a tone and domain. It cannot repair a vague request.

Weak version:

You are an expert. Make a promotion strategy.

Working version:

Below is a description of a B2B SaaS product for CFOs. Prepare a 90-day promotion plan for a $3,000 budget and a team consisting of the founder and one marketer. For five channels, give the hypothesis, test cost, success metric, and risk; do not invent market figures without a source.

The second request provides inputs and a way to review the result. The role label is secondary.

Practical working habits

Choose the instruction language through testing. Some models follow structure more consistently in English. When your own evaluations confirm that behavior, write the working instruction in English and specify the response language separately.

Ask AI to draft the prompt. In a separate chat, describe the task and ask the model to collect missing inputs before preparing a request for another model.

Work in iterations. Treat the first response as a draft. You can then request sources, shorten the text, change the tone, or separate facts from assumptions.

Keep tasks in separate chats. A contract, advertising copy, and a strategy in one history create conflicting context.

Use voice input for long briefs. It saves time when you need to explain what happened, why it matters, and which constraints apply.

Anthropic lists several Claude Console tools for creating and refining requests:

“The Claude Console also offers prompting tools—prompt generator, templates and variables, and prompt improver—to help you build and refine prompts quickly.” — Anthropic Claude Docs

Source: Anthropic Claude Docs, Prompt engineering overview

Templates and generators can speed up preparation. The user still needs to understand the task and its acceptance criteria.

How to review the answer

A confident tone does not prove accuracy. For professional work, ask the model to provide:

  • a separation of facts and assumptions;
  • sources for external data;
  • risks and weaknesses in the answer;
  • an alternative option;
  • a list of missing inputs.

Contracts, financial calculations, legal and medical material, and public analysis require review by a qualified person. AI can prepare the draft, while responsibility remains with the human reviewer.

A reusable template

Context:
[Company, product, audience, and situation]

Task:
[The specific action]

Input data:
[Facts, documents, links, and constraints]

Output format:
[List / table / email / plan / JSON]

Quality criteria:
- separate facts from assumptions
- do not invent numbers
- mention risks
- ask questions if data is missing

Adapt the template to the task. Open examples are available in the Anthropic Prompt Library and Microsoft documentation.

Copying an example word for word is rarely useful. Examine its context, format, constraints, examples, and review criteria, then build the request your task needs.

A prompt creates controlled input for the system. A precise task and clear conditions make the first draft useful sooner.