The art of prompting: how to write a better prompt in 10 steps

An AI model can do a lot. Sometimes even an alarming amount. But it doesn’t automatically know what matters to your target audience, what claims you’re allowed to use, or how your organization communicates.

That’s where prompting comes in.

With a good prompt, you give AI enough direction to deliver a useful result, not by spelling out every step, but by making it clear what you want to achieve, what information is relevant, and what criteria the output must meet.

This is especially important now that models like GPT-5.6, Claude Fable 5, and Claude Opus 5 can handle more complex tasks and follow instructions more closely. As a result, the best prompt isn’t automatically the longest one. In fact, a prompt full of repetition, mandatory steps, and conflicting rules can actually worsen the result. For GPT-5.6, OpenAI recommends stating instructions only once and, above all, providing context, hard constraints, boundaries, and success criteria. Anthropic warns that prompts designed for earlier models can sometimes be too prescriptive for Claude Fable 5.

We’ll show you how to write a better content prompt in ten steps.

 

The 10 building blocks of a good prompt

1. Task, what should the AI do?
2. Goal and context, what should the output achieve, and why?
3. Target audience, who is the output intended for?
4. Tone of voice, how should it come across?
5. Style, how should the sentences and words be written?
6. Output, what exactly should the AI produce?
7. Knowledge, what information can the AI use?
8. Process, are certain intermediate steps required?
9. Conditions, what requirements must the output meet, and what is the AI not allowed to do?
10. Quality criteria, when is the result considered successful?

Not every prompt needs all ten components.

For a short social media post, the task, target audience, tone of voice, and output are often sufficient. For an in-depth research article, sources, conditions, and quality criteria become more important. And when AI also uses tools or can perform actions, clear boundaries are not a luxury but a necessity.

So use these steps as building blocks, not as a bureaucratic checklist. We’re writing a prompt, not a bid for a new highway.

Discover the step-by-step guide

 

What is prompting?

Prompting is the process of providing instructions and information to an AI model to achieve a desired result.

This can be a simple question:

"Summarize this text in five bullet points."

But a prompt can also be a detailed task, such as writing a landing page, analyzing research data, or performing multiple steps within a business process.

You can enter a prompt directly into ChatGPT or Claude. Prompts can also be part of an AI assistant, an agent, an API integration, or a platform like Sterc.ONE.

Think of prompting as training a smart new colleague. That colleague has a lot of knowledge and can work quickly, but doesn’t yet know your organization, target audience, agreements, or preferences.

You don’t need to explain to that colleague how every thought should be formed. You do, however, need to make it clear:

• what the task is;
• why that task is important;
• what information should be used;
• what requirements the result must meet;
• what the boundaries are.

That, in essence, is what a good prompt does.

Different types of prompts

For content and marketing, three types of tasks are most common:

• Writing: AI creates new content, such as blogs, social media posts, emails, ads, job postings, and landing pages.
  For a writing task, the target audience, tone of voice, style, and output are particularly important.
• Analysis: AI evaluates existing information and derives insights from it. Examples include an SEO analysis, a content audit, competitive research, evaluating a landing page, or analyzing customer feedback.
  For an analysis task, criteria, sources, supporting evidence, and desired depth are more important than tone of voice.
• Analysis and writing: AI first analyzes the input and then creates an improved version. For example: analyzing and rewriting an existing web page, translating customer interviews into a landing page, analyzing search results and then creating a content brief, or evaluating and optimizing a job posting for the ideal candidate.

That initial choice is important. A writing task requires tone of voice, style, and format. An analysis task, on the other hand, requires criteria, sources, and rigorous validation. If you lump everything together, you’ll end up with output that’s a jumble. It might be entertaining, but it’s rarely useful.

What is content prompting?

Content prompting is the targeted use of AI to create, evaluate, and improve content.

It’s not about:

“Just write a blog post about AI.”

It’s about guiding AI based on the desired goal, the target audience, brand guidelines, sources, structure, and conditions.

At Sterc, we don’t just use AI to produce text faster. We use AI to streamline content processes. This lets you move more quickly from input to draft, maintain a more consistent tone of voice, and tailor your content more precisely to the target audience and the customer journey.

Content prompting helps you, among other things, to:

• tailor content to the target audience and their level of knowledge;
• maintain a consistent tone of voice;
• utilize sources and your own company knowledge;
• choose the right format;
• establish SEO, branding, and legal requirements;
• create multiple relevant versions more quickly;
• spend less time on generic first drafts.

 

Why is a good prompt so important?

An AI model makes decisions based on your prompt and the available context.

If you don’t specify who the text is intended for, the model will choose a level on its own. If you don’t state a goal, it will determine for itself which arguments are important. If you don’t provide reliable sources, general or incomplete information may end up in the text.

That doesn’t have to be a total disaster. Modern models are getting better and better at understanding what a user likely means. For example, GPT-5.6 is better at inferring the underlying intent and desired amount of work, so you don’t have to specify every single step of the process by default. OpenAI does, however, recommend continuing to specify relevant domain context, hard constraints, boundaries, and success criteria.

Let’s look at a practical example:

❌ Vague: “Write a text about marketing.”

✅ Specific: “Write a Dutch LinkedIn post for small business owners about three ways they can use their Google Ads budget more effectively. Use a maximum of 150 words, provide one concrete example, and end with a question that encourages comments. Write in a practical and direct style, without making any guarantees about results.”

The second prompt isn’t better simply because it’s longer. It’s better because the information directly influences the outcome.

 

What sets modern AI models apart

With older AI models, detailed prompts and fixed step-by-step plans were often necessary to get usable output. Newer models can process more context, better recognize intentions, and independently carry out more complex tasks.

And that’s exactly what changes the way you craft prompts.

Be less prescriptive, set clearer boundaries

GPT-5.6 is better at deducing which approach suits a task. OpenAI therefore recommends focusing on specifying the desired outcome, relevant context, conditions, required justification, and success criteria. You don’t need to routinely ask the model to think harder or generate multiple answers.

Claude Fable 5 follows short instructions more closely than earlier Claude models. Anthropic even recommends reevaluating older prompts, as too many instructions and scaffolding can reduce quality.

The conclusion is simple: Tell the AI what a good result looks like. Don’t unnecessarily prescribe how it should get there internally.

Instructions are followed more literally

Stronger instruction following also means that contradictions can have a more significant impact.

Asking for a text that is both very detailed and no more than 200 words? Then the model must decide which instruction takes precedence.

Are you asking for a critical analysis, but also stating that the existing text is excellent? Then you’re already steering the conclusion before the analysis even begins.

Therefore, provide each important instruction only once and place strict requirements in a logical spot. OpenAI explicitly recommends this for GPT-5.6.

Length varies by model

Not every model responds the same way to general instructions like “be concise.”

By default, GPT-5.6 is more concise than its predecessor. An additional instruction to keep the response very short can therefore cause important nuances to be lost. OpenAI recommends specifying, for short output, which information must be retained and what can be omitted first.

Claude Opus 5, on the other hand, produces relatively detailed responses and documents. Anthropic therefore recommends explicitly tailoring the desired length to the task and excluding unnecessary summaries or filler.

So “Keep it short” is sometimes too simplistic.

Does the language of your prompt matter?

Less than is often assumed.

If you want Dutch output, you can easily write the prompt in Dutch. English can be useful for technical settings, programming languages, or standard API terms, but it’s not a secret shortcut to better answers.

Consistency is more important.

For example, you can easily use English labels and write the content in Dutch:

"Task: write a blog intro
Goal: raise awareness among entrepreneurs about the impact of website speed
Audience: SME entrepreneurs with limited technical knowledge
Tone of voice: down-to-earth, practical, and direct
Language: Dutch
Output: maximum 150 words"

No prompt poetry. Just clear.

 

10 steps to a better prompt

1. Task

Don’t start with who the AI should be, but with what the AI should do.

Should the model write, analyze, compare, summarize, or improve something, or should it analyze first and then write a new version?

A clear task prevents the model from interpreting your instructions differently than you intended.

Example for a writing task:

"Task: Write a blog post about why website speed is important for small and medium-sized businesses."

Or for an analysis task:

"Task: analyze this landing page for SEO, clarity, and conversion effectiveness. For each section, identify what’s working well and what needs improvement."

Should you assign a role to your AI?

Assigning a role can be useful when a specific perspective, area of expertise, or behavior is important.

For example:

"Evaluate the text from the perspective of an experienced SEO specialist."

Or:

"Write as a consultant who makes technical topics understandable for business owners."

Don’t use a role just for show. “You’re the best marketer in the world” sounds nice, but it doesn’t say much about what the model is actually supposed to do.

 

2. Purpose and context

Describe why you’re having the content created and in what situation it will be used.

The purpose helps AI make content-related decisions. The context clarifies the reason behind the content and where the text needs to fit.

Do you want to inform, persuade, motivate, compare, or build trust? Will the text be used on a new landing page, in a campaign, or as a follow-up after a workshop?

You don’t need to paste your entire company archive into the prompt. Just include context that influences the content, angle, or desired action.

Example:

"Goal and context: raise awareness among small and medium-sized business owners about the impact of website speed on user experience, searchability, and conversion. The article will be published on an online agency’s website and should encourage readers to have their website performance analyzed."

Explaining why an instruction is important can also help a model make better decisions within the task. Anthropic therefore recommends including not only the rule but also the underlying reason for relevant instructions.

 

3. Target audience

Describe who the output is intended for.

Specify, where relevant:

• position or role;
• level of knowledge;
• primary need;
• familiarity with the subject;
• doubts or resistance;
• stage in the customer journey.

“For entrepreneurs” is often too broad. A self-employed person just starting out has different questions than a marketing manager at an organization with a hundred employees.

Example:

"Target audience: SMB entrepreneurs and marketing managers with limited technical knowledge. They know their website is important, but they still mainly view speed as a technical issue for developers. Above all, they want to know what business impact speed has and when action is needed."

 

4. Tone of voice

Describe how the sender should come across to the reader.

Think of words like down-to-earth, critical, reassuring, enthusiastic, advisory, motivating, or opinionated.

Terms like “professional,” “modern,” and “fresh” are often too general, so show how the tone sounds in practice.

Example:

"Tone of voice: down-to-earth, practical, and slightly opinionated. Address the reader directly and be persuasive without coming across as patronizing or overly enthusiastic. Write as if an experienced marketer were giving an entrepreneur a wake-up call in an approachable way."

A short example sentence can sometimes convey the desired tone better than a laundry list of separate adjectives.

 

5. Style

Describe how the text should be written at the sentence and word level.

Consider language level, sentence length, active or passive voice, form of address, word choice, use of technical terms, and words you want to avoid.

Tone of voice and style are not the same thing.

The tone of voice determines how the text feels. The style determines how that is reflected in the wording.

Example:

"Style: Write in Dutch at the B1–B2 level. Use mostly short, active sentences and address the reader as ‘you.’ Choose common Dutch words whenever possible. Use technical terms only when they’re necessary for the content, and explain them briefly when you do. Avoid long-winded sentences, abstract phrasing, and unnecessary English.”

You can support the desired style by writing your prompt clearly yourself. But don’t turn your prompt into a miniature blog. Concrete instructions and relevant examples are more important.

 

6. Output

Specify exactly what the AI should produce.

Specify the following where necessary:

• the content type;
• the desired length;
• the structure;
• required components;
• the desired level of detail;
• the format in which the result must be delivered.

A blog post, LinkedIn post, email, and analysis each require a different structure.

If there’s a strict length limit, don’t just provide a word count. Also specify what must be kept if the text needs to be shortened.

Example:

"Output: Write a blog post of approximately 900 words. Start with a short, relatable introduction. Use clear H2 headings, short paragraphs, and bullets only when they make the information easier to follow. Briefly explain Core Web Vitals, use practical examples, provide concrete tips, and end with a specific CTA. If space is limited, first remove repetition and general background information, not the most important arguments or examples.”

 

7. Knowledge and sources

Specify what information the AI should use as a basis.

Consider:

• provided documents;
• existing web content;
• interviews;
• research results;
• brand documentation;
• product information;
• reliable external sources.

Make it clear which sources are authoritative and whether AI is allowed to look up or use additional information.

Also specify what the model should do when information is missing, unclear, or contradictory. In such cases, don't let the AI guess. Have it identify uncertainties or indicate what information is still needed.

Do you want verifiable claims? If so, specify whether links, footnotes, or a list of sources are required.

Example:

"Knowledge and sources: Use the provided information about Core Web Vitals and MODX as your primary source. Use additional external sources only if they are up-to-date, reliable, and relevant. Don't include figures or claims you can't substantiate. Indicate when information is missing or sources contradict each other. Add a source citation to relevant figures."

Tip for prompts with a lot of source material

Are you working with multiple long documents? If so, keep that information clearly separate from the assignment.

You can use regular subheadings, but for complex prompts, XML tags can also be helpful:

<sources></sources>
Sources and background information

<instructions></instructions>
The task and conditions

Anthropic recommends using XML tags especially for prompts where instructions, context, examples, and variable input may be intermixed. For very long documents, Anthropic recommends placing the source material at the top and providing the final question afterward.

Of course, you don’t need all that for a short LinkedIn post.

 

8. Method

Only describe a procedure when the order affects the result.

A fixed order can be useful for:

• extensive analyses;
• research articles;
• content based on multiple sources;
• assignments with an interim assessment;
• processes that require prior approval.

Above all, specify what the AI should produce and in what order. Avoid unnecessarily dictating how the model should think internally.

Instructions such as “think step by step,” “check everything three times,” or “come up with five answers first” do not automatically improve the outcome. Furthermore, Claude Fable 5 should not be asked to reproduce its entire internal reasoning in the final output.

Example:

"Procedure:

1. First, analyze the provided sources.
2. Then create a concise article structure with H2 headings.
3. Next, write the entire article.
4. Base factual claims on the provided sources.

Submit only the final article."

For a simple assignment, it can be much shorter:

"Analyze the input and then write the final text right away. Do not create a separate outline."

 

9. Requirements

Specify the strict requirements the output must meet and where the assignment ends.

For a standard content prompt, these might include SEO guidelines, brand guidelines, required topics, prohibited claims, privacy rules, legal terms, words to avoid, and maximum length.

State each rule only once.

Example:

"Requirements:

• Incorporate the main keyword ‘make your website faster’ naturally into the title, introduction, and at least one subheading.
• Use only numbers and claims supported by the sources.
• Don't make any guarantees about rankings, revenue, or conversion rates.
• Do not mention competitors.
• Avoid words like ‘revolutionary,’ ‘groundbreaking,’ and ‘game-changer.’
• Do not write a technical guide for developers.
• The assignment is complete once the article, meta title, and meta description have been delivered."

Limitations for assistants and agents

Can AI not only provide advice but also edit files, send emails, or publish content? If so, please also describe:

• which actions the AI is permitted to perform independently;
• what falls outside the scope of the assignment;
• when permission is required;
• when the task is complete.

GPT-5.6 can work proactively and persistently within multi-step tasks, so OpenAI recommends setting explicit approval boundaries for external, destructive, costly, or scope-expanding actions.

Example for an AI assistant with CMS access:

"Analyze the existing page and create an improved draft version. Don't publish anything without explicit approval. Don't modify URLs, forms, tracking codes, or technical settings. Request approval whenever a change could affect SEO, functionality, or existing campaigns."

 

10. Quality criteria

Describe what a successful output must meet. Consider completeness, relevance to the target audience, required arguments, use of sources, technical accuracy, readability, CTA, and desired length.

A general instruction like “check your answer” is often too vague. Quality criteria give the model a clear picture of the desired end result.

Example:

“Quality criteria: the text is tailored to small and medium-sized business owners, clearly illustrates the business impact of website speed, uses only substantiated claims, incorporates the main keyword naturally, and ends with a specific CTA.”

An additional mandatory self-check isn’t necessary for every task or every model.

Claude Opus 5 already performs rigorous verification of its work by default. Anthropic recommends removing generic verification steps from older prompts, as these can lead to excessive checking and additional token usage. For long, autonomous tasks with Claude Fable 5, however, a separate verification step or an independent reviewer can indeed be valuable.

So, distinguish between:

quality criteria: almost always useful;
an additional verification process: use only when the task or risk justifies it.

 

Example prompt

The prompt below uses all ten building blocks. For a shorter assignment, you can omit certain parts.

Task
Write a blog post about why website speed is important for small and medium-sized businesses.

Goal and context
Raise awareness among business owners about the impact of website speed on user experience, search visibility, and conversion rates. The article will be published on an online agency’s website and should encourage readers to have their website performance analyzed.

Target audience
SME business owners and marketing managers with limited technical knowledge. They know their website is important, but still tend to view speed primarily as a technical issue for developers.

Tone of voice
Down-to-earth, practical, and slightly cheeky. Address the reader directly and be persuasive without sounding patronizing or overly enthusiastic.

Style
Write in Dutch at the B1–B2 level. Use mostly short, active sentences and address the reader informally with “je.” Use technical terms only when necessary for the content and explain them briefly. Avoid long-winded introductions and abstract marketing jargon.

Output
Write a blog post of approximately 900 words. Start with a short, relatable introduction. Use clear H2 headings and short paragraphs. Briefly explain Core Web Vitals, provide practical examples and concrete tips, and conclude with a specific CTA. Also include a meta title and meta description.

Knowledge
Use the provided information about Core Web Vitals and MODX as your primary source. Use additional external sources only if they are up-to-date, reliable, and relevant. Do not fabricate figures, case studies, or claims. Mention it if essential information is missing.

Workflow
First, analyze the sources and then create a logical article structure. Next, write the entire article. Submit only the final version.

Requirements
Incorporate the main keyword “make your website faster” naturally into the title, introduction, and at least one subheading. Do not make any guarantees regarding rankings, revenue, or conversion rates. Do not mention competitors. Do not write a technical guide for developers.

Quality criteria
The text is tailored to the target audience, clearly illustrates the business impact of website speed, uses only substantiated claims, incorporates the main keyword naturally, and concludes with a specific CTA.

Is this a detailed prompt? Absolutely.

Is such a detailed prompt necessary for every piece of text? Absolutely not.

That’s exactly why you use the ten steps as building blocks.

 

Use examples when description alone isn’t enough

Do you want a very specific tone of voice, structure, or phrasing? Then add a good example.

Examples are a reliable way to guide Claude on format, tone, and structure. They do need to be relevant and representative, though. After all, a bad example also teaches the model very efficiently what not to do.

When providing an example, specify what the AI should adopt:

"Use the text below as a style example. Adopt the direct tone, short paragraphs, and down-to-earth humor. Do not copy any sentences, examples, or content."

Don’t automatically add five examples to every prompt. Examples are useful when they clarify a specific brand guideline or solve a demonstrable problem.

 

Examples of good and bad prompts

Every department within your company can use AI tools like ChatGPT and Claude in different ways. Below are a few specific examples:

Marketing

❌ Too vague: “Write a landing page for our marketing tool.”

The AI doesn’t know who the page is for, what the main problem is, or what action the visitor should take.

✅ More specific: “Write a Dutch landing page for SMB marketers who are struggling to get a handle on their Google Ads results. Convince them to request a free demo of our marketing tool. Use a direct, practical tone of voice and write approximately 450 words. Structure the page with a problem-focused H1, a short introduction, three concrete benefits, a brief, substantiated customer case study, and a clear demo CTA. Use only the provided product information and do not invent any features, percentages, or customer results."

 

Sales

❌ Too vague: “Create a presentation about our software.”

The model doesn’t know whether you need a pitch, a slide structure, a script, or a product sheet.

✅ More specific:
"Write the content for a Dutch-language sales presentation on transportation management software. The presentation is intended for logistics managers at medium-sized companies and should lead to a demo request. Address common challenges such as error-prone scheduling, limited visibility, and high transportation costs. Create a structure of up to eight slides, with a title and up to three key points per slide. Use only substantiated product benefits and do not fabricate results or customer case studies.”

 

HR

❌ Too vague: “Write a job posting for a software developer.”

Seniority, technical stack, employment terms, and target audience are missing. The risk of ending up with “Are you an enthusiastic team player with a passion for code?” is now dangerously high.

✅ More specific:
"Write a Dutch job posting for a senior software developer with at least five years of experience. Engage candidates by highlighting technical challenges, autonomy, and the opportunity to influence architectural decisions. Use a professional yet personal tone of voice. Structure the text with a relatable opening, a description of the role, five key responsibilities, five concrete benefits, and a clear call-to-action (CTA) for applicants. Use only the information provided. Note any missing information regarding salary, tech stack, or work style as a point of attention, and do not fill in these details yourself."

 

Tip!
Have AI critically review your job posting from the perspective of your ideal candidate. First, describe who that candidate is and what they look for. Then, have AI identify the biggest deal-breakers and rewrite the job posting.

 

7 rules for better prompts