Prompting still matters
The progress in generative AI over the last few years has been tremendous, and modern language models need far less hand-holding. You do not need to learn a secret vocabulary or fill every prompt with complicated instructions. That does not mean prompting is any less important. Good prompting still matters.
A vague request will often produce a generic answer. A model given useful context, examples and clear expectations can generate results that are genuinely useful — and, just as importantly, that match your tone and sound like your brand.
The goal is not to delegate your thinking to an LLM. It is to use the model to improve your work: explore ideas, organise information, and reduce repetitive tasks. All of this can be done with a simple prompt, but the results will fall well short of what is possible, and the chance of a mistake is much higher.
Fortunately, you can quickly learn how to elevate your LLM instructions. Here are six prompting techniques that can help.
1. Zero-shot prompting
Let's start with the technique you probably use most often. Zero-shot prompting means asking the model to complete a task without showing it any examples. This is the most common way people use tools such as ChatGPT. It works well when the task is simple or the expected outcome is easy to explain.
For example: write a LinkedIn post announcing a practical AI workshop for small business owners, keep it below 150 words, and focus on the problems the workshop will help them solve.
This prompt does not include an example post, but it still defines the audience, format, length and main message. A zero-shot prompt does not have to be simple — it can include detailed context, constraints and outcome criteria without including specific examples.
This type of prompt is often useful as a starting point to quickly assess whether the model's response is good enough for you. Use it when you need a quick draft, a summary, a list of ideas, or an answer to a clearly defined question. Analysing the answer from a zero-shot prompt can give a good indication of what type of information is needed in the prompt to get a much better result.
2. Few-shot prompting
Few-shot (or sometimes many-shot) prompting gives the model one or more examples to analyse before asking it to complete the task. These examples help the model understand patterns such as the structure, tone or level of detail you expect.
There are plenty of tasks where you use a fairly consistent format or structure. Imagine you regularly publish LinkedIn posts about your company. Rather than describing the company's writing and brand style, including a few past posts can generate a much better outcome: give the model two examples of previous posts, and ask it to write a new post announcing the workshop, matching the tone, sentence length and level of detail used in the examples.
You can also include examples of what to avoid — such as exaggerated claims like "AI will transform your business" or "this changes everything."
It is important to remember that the quality of the examples matters a lot. If your examples are unclear, inconsistent, badly written or contain mistakes, the model may follow them and generate similar issues in the final response.
Repeated tasks are especially useful to combine with few-shot prompting. Adding a few examples can greatly improve an LLM's responses in areas like classifying customer messages, writing product descriptions, or creating content in a consistent format.
Examples can also reduce variation, which is especially visible in software development. However, this should not be your only safeguard. Structured outputs, validation and error handling are still needed when another part of an application depends on the response from an LLM.
3. Structured reasoning
A lot of tasks require more than a direct answer. The model may need to identify missing information, compare options, or work through several parts of a problem before providing an answer. In cases like these, following a clear reasoning process or workflow before producing the final result can be a better option.
For example: ask the model to help plan a LinkedIn campaign for the AI workshop by first identifying any important information missing from the brief, then suggesting three possible angles for the campaign, comparing their strengths and weaknesses, and finally recommending one angle with a brief explanation of the choice.
The goal is to make the model slow down and analyse the task from various angles. If you already have a defined approach for a task, you can ask the model to use a specific framework. This is more useful than asking the model to "think harder" — it gives the model a clear structure, and you can review the answer against a defined outcome.
It is worth asking for assumptions, conclusions and checks, rather than a long transcript of the internal reasoning. The goal is to make the answer easier to assess, not simply longer. This matters especially when the answer is used further in the workflow, since long transcripts can take up a significant portion of the context window.
4. Ensembling
Ensembling means generating several answers and comparing them instead of relying on the first response — by running the same prompt several times, changing the structure or specific words, asking different models, or giving each response a different role. This technique comes from one of the most popular AI methods of combining two or more algorithms to generate better overall results.
For example: ask the model to create four opening hooks for a LinkedIn post promoting an AI workshop, one focused on wasted time, one on risk, one challenging a common assumption, and one starting with a practical example, then rank them based on clarity and relevance to small business owners.
An interesting option is to generate several versions of the answer and ask the model to identify the strongest elements from each one. This is useful because responses can vary due to the nondeterministic nature of LLMs — one version may have a strong opening but a weak conclusion, while another may explain the offer clearly but sound too formal.
Ensembling helps you compare options, find areas of agreement and notice unusual answers. It is particularly useful for tasks where there is not one obvious correct response, and where responses might vary greatly. Beyond changing the prompt itself, adjusting options like temperature and top-p can also produce distinctly different answers.
The clear downside of ensembling is cost and time. Generating five versions of every routine task, like an email draft, is rarely necessary, especially when the template is already well defined. Ensembling should be used when the quality of the decision justifies the extra work, and when the nondeterministic nature of the models turns into an advantage rather than noise.
5. Self-critique and revision
Models can review not only a user's messages but, just as importantly, their own. A model can review its own answer against a set of criteria, then identify the shortcomings and generate an improved version.
For example: ask the model to review the LinkedIn post it just wrote, checking whether the opening is specific, the post explains who the workshop is for, any claims sound exaggerated, and the call to action is clear — then to briefly list the main problems before rewriting the post.
Just as with the examples in few-shot prompting, the criteria matter. Asking "can you make this better?" leaves far too much room for interpretation, when your goal is to narrow the scope and guide the model in a specific direction. A focused review can improve clarity, structure and consistency, and can also catch missing information or unsupported claims.
Keep in mind that self-critique is not the same as factual verification. If the model invented a figure or statistic, asking it to review the answer may not correct the problem — the model might simply repeat the same mistake with more confidence. Important claims should still be checked against trusted documents, reliable sources, or data from your own systems.
6. Decomposition
Working on complex tasks often requires creating multilayered prompts. These often contain several tasks or steps hidden inside one request. A model may complete each part, but the quality can still vary between them. Especially when the process is long, the model might lose focus and overlook earlier constraints, or rush through the final stages.
For example, asking a model to research the market, define the audience, create a campaign strategy, write five posts and prepare a reporting plan in one request hides six different jobs inside a single instruction.
Decomposition means breaking the task into smaller steps that follow a defined workflow: define the target audience, identify the audience's main problems, choose the campaign message, create the content plan, write the individual posts, and review them against the original goal.
Each stage produces a partial output that can be checked before moving to the next step. This structure helps the model, but it is also useful to the person running the process — for example, a person may notice that the campaign is based on the wrong audience before spending time writing ten posts.
Decomposition works particularly well for detailed, complicated tasks like research, content production, data analysis, and AI workflows involving several tools. The output of a decomposition prompt is often used for further processing — for example, data that populates a CRM record and is then used by other workflows.
It can be overused. A simple automated email does not need a six-stage process. Break work into steps when the task is genuinely complex, not because longer prompts look more advanced.
Combining prompting techniques
These techniques are often most useful when combined. For example, you could decompose a campaign into research, planning and writing stages; use a few strong examples during the writing stage; generate several campaign hooks through ensembling; and review the final content against a short checklist.
The best combination depends on the task. There is no reason to use every technique in every prompt. Start with the simplest approach, and add more structure only when the first result is inconsistent, incomplete or difficult to evaluate. Add as much to the prompt as you need, but as little as necessary, to avoid clogging the context window.
Prompting is only part of the process
A good prompt improves the chance of receiving a useful answer. The other side of the coin is the nondeterministic nature of large language models. Even the best prompt does not guarantee that the answer will be correct.
The goal of prompting is to narrow the scope and steer the model in the right direction, to limit the chance of an incorrect answer. When AI supports an important business process, prompting should sit alongside source checking, clear output formats, automated validation, human review, and testing against real examples.
This matters even more when an AI response feeds directly into software development, customer communication or business logic.
Final thought
Good prompting is not about finding magic words, at least not anymore. It is about giving the model the information it needs, showing what good work looks like, and creating a repeatable process to evaluate the result.
Choose one real task you complete regularly. Test it first with a clear zero-shot prompt. If the result is inconsistent, add an example, break the work into stages, or introduce a review step.
This is just the beginning of prompting techniques. A detailed research taxonomy of techniques and their applications is worth reading if you want to go further.
Which of these techniques would you like us to demonstrate using a real business workflow? Let us know if your processes need to be automated or you need help implementing AI in your business.