Prompt engineering
Prompt Engineering simply means asking good questions to the AI. A prompt is your input with your question or request.
Prompt engineering involves devising targeted questions and instructions to help AI models provide the right answers during chats. With better prompts, you get better results.
Together with the specifications of the chosen model, this determines the quality of the answer.
Elements of a prompt (question)
To get better results, you should consider the following points:
- Goal: what exactly do you want from the language model?
- Role: what role does the language model have? (e.g., 'Marketing Analyst', 'Legal Advisor', 'Project Manager')
- Context: background information about the request (e.g., business objectives, project details, market information)
- Style: what style do you want to use? (e.g., formal, persuasive, concise)
- Output: what format do you want? (e.g., a report, email, presentation outline, code snippet)
The goal
Precisely describing the goal of your question is essential and cannot be omitted.
What do you want to know?
- Fact-based questions are possible, but be aware of limitations (e.g., lack of up-to-date knowledge)
- Just like with Google, you won't get exactly the same answer if you ask the question multiple times
- The advantage is the “chat” function so you can ask follow-up questions
Example prompt with a clear goal:
Provide a description of the main challenges in implementing a new AI platform in a medium-sized company.
Implementing a new AI platform in a medium-sized company involves several challenges. Some key challenges are:
- Resistance to change: Employees are accustomed to existing processes and systems. A new AI platform requires adjustment and may cause resistance, especially if the impact on workflows is unclear.
- Data quality and availability: AI platforms heavily depend on data. Collecting, cleaning, and structuring data for the new platform can be complex and time-consuming.
- Integration with existing systems: The AI platform often needs to be integrated with other business applications (ERP, data sources, BI tools). This can pose technical challenges.
- Costs: Besides platform licensing costs, there are costs for implementation, training, data infrastructure, and possibly specialized personnel.
- Training and adoption: Employees must be trained in using the new platform and interpreting the output. Without good training and continuous support, adoption rates may remain low.
- Choosing the right platform and use cases: There are many AI platforms and technologies. Selecting a platform and defining relevant use cases that fit the specific needs and strategic goals of the company is crucial.
- Ethical considerations and compliance: Depending on the AI platform's application, there may be ethical issues and compliance requirements (e.g., GDPR) that must be carefully addressed.
- Project management: Successful implementation requires solid project management, clear objectives, and involvement of stakeholders from various departments.
Adding role and context
Adding a role and context results in better and more specific answers.
Example of the same question but with role and context:
As a project manager responsible for implementing a new AI platform for data analysis and process optimization at a B2B service provider with 150 employees, provide a description of the main challenges and possible mitigation strategies.
As a project manager for implementing a new AI platform for data analysis and process optimization at a B2B service provider with 150 employees, I see the following key challenges and possible mitigation strategies:
- Resistance to change & User adoption:
- Challenge: Employees, especially those whose tasks are directly affected, may be reluctant to abandon familiar workflows and embrace new AI-driven insights.
- Mitigation: Engage key users early. Clearly communicate benefits (efficiency, better decision-making, new insights). Provide extensive role-based training and continuous support. Appoint 'AI champions' within departments.
- Data quality, governance, and migration:
- Challenge: The AI platform's effectiveness depends on high-quality, well-managed data. Data from existing systems is often inconsistent, incomplete, or incorrect.
- Mitigation: Start with a thorough data audit, define a data governance framework, and perform data cleansing before integration/migration. Define clear data models and flows. Conduct tests and carefully validate data.
- Integration with existing IT infrastructure:
- Challenge: Connections with existing data sources, ERP, CRM, and other business systems must work seamlessly and securely.
- Mitigation: Identify all required integrations early. Choose an AI platform with robust API capabilities or standard connectors. Test integrations thoroughly.
- Scope creep & Budget overruns:
- Challenge: During the project, new requests for AI applications may arise that broaden the scope and increase costs.
- Mitigation: Define a clear, bounded scope with well-defined use cases (MVP - Minimum Viable Product) for the first phase. Apply a strict change management process. Maintain a realistic project schedule and budgeting, including a buffer.
- Selecting the right technology and implementation partner:
- Challenge: The AI market is complex. Choosing the wrong technology or an inexperienced partner can cause project failure.
- Mitigation: Conduct thorough research on suitable AI technologies for the defined use cases. Select a partner with proven experience in your industry and with the chosen technologies. Request references.
- Lack of AI knowledge and skills within the organization:
- Challenge: There may be insufficient internal expertise to effectively use, maintain, or further develop the AI platform.
- Mitigation: Invest in training and education. Consider hiring specialized personnel or work closely with the implementation partner for knowledge transfer. Foster a culture of continuous learning.
- Ethical considerations and explainability (Explainable AI):
- Challenge: Decisions made or supported by AI must be transparent and explainable, especially if they impact customers or employees. Bias in data can lead to unfair outcomes.
- Mitigation: Establish ethical guidelines for AI use. Prefer interpretable models where possible. Implement processes to detect and mitigate bias. Ensure human oversight on critical AI decisions.
By proactively addressing these challenges, we increase the chances of a successful AI platform implementation that truly contributes to business objectives.
Techniques
Below we describe several techniques you can use for more advanced prompts.
Technique 1 - providing examples
To ensure the language model uses the style you want, provide one or two examples.
This method is called “one & few-shot prompting”
Write a draft email to customers about an upcoming planned system update. Use our internal policy for customer communication (include relevant points from policy). Use this recent newsletter as a style guide, but do not include substantive information: (insert newsletter text)
The full prompt is as follows:
Internal policy for customer communication:
- Always be professional and polite.
- Provide timely updates about important changes or outages.
- Offer clear and concise information.
- Provide a solution or compensation if applicable.
Use this newsletter for style, but do not use information from this newsletter. Dear customer,
We would like to inform you about an upcoming system update that may affect our services.
Date and time of update: [date and time] Expected duration of interruption: [duration] Affected services: [services]
During this period, our services may be temporarily interrupted. We will do our best to minimize inconvenience and restore services as soon as possible.
Below is the language model's answer to the detailed prompt with examples
Dear customer,
We would like to inform you about an upcoming system update that may affect our services.
Date and time of update: [date and time] Expected duration of interruption: [duration] Affected services: [services]
During this period, our services may be temporarily interrupted. We will do our best to minimize inconvenience and restore services as soon as possible.
If you have any questions, please feel free to contact us.
Kind regards, [Your Name] [Your Position] [Company Name] [Contact Information]
Technique 2 - providing a template
To ensure the language model uses a specific format, provide a template and an example. This ensures the language model structures or formats its answer in a certain way. You can think of very different formats useful in situations such as:
- csv output for importing into an Excel sheet or data analysis tool
- json output for API integrations or configuration files
- Agenda setup for a project meeting
Example prompt: Generate a list of three potential marketing channels for a new SaaS product aimed at SMEs. Format your answer as follows: Channel: Target audience reach: Potential ROI: Required resources:
For example: Channel: LinkedIn Advertising Target audience reach: Specific job titles, industries, company size Potential ROI: High, if well targeted Required resources: Advertising budget, content creation, campaign management
This method is called “output templates”
Below is the language model's answer to the prompt with an output template
Channel: Content Marketing (Blog & SEO) Target audience reach: Companies actively searching for solutions to their problems, organic traffic Potential ROI: High in the long term, builds authority Required resources: Content writers, SEO specialist, time for keyword research and content promotion
Channel: Google Ads (Search) Target audience reach: Companies actively searching for specific keywords related to the SaaS product Potential ROI: Medium to high, depending on competition and CPC Required resources: Advertising budget, keyword research, landing page optimization, campaign management
Channel: Email Marketing (Lead Nurturing) Target audience reach: Leads collected via website, webinars, or other channels Potential ROI: High, effective for converting leads to customers Required resources: Email marketing platform, content for email sequences, segmentation skills
Technique 3 - giving a hint
To ensure the language model starts and writes in a certain way, provide a hint. Here we write the hint at the end of the prompt.
Example prompt: Can you summarize this market research report for me? (paste the relevant sections of the report). Start your answer with: The main findings from this market research are:
- PASTE THE TEXT OF THE REPORT
This method is called “cues and hints” This example uses a fictional report.
Below is the language model's answer to the prompt with a hint
The main findings from this market research are:
- Demand for cloud-based project management software within SMEs has increased by 35% over the past two years.
- Key decision factors for SMEs are ease of use, integration possibilities with existing tools, and pricing.
- Competitor A has a market share of 25%, followed by Competitor B with 18%. There is room for new players focusing on specific niches or offering a superior user experience.
- The biggest challenges for SMEs in adopting new software are the initial investment and the time needed for employee training.
- The market is expected to continue growing annually by 15% over the next five years.
Technique 4 - splitting content from instructions
Sometimes the prompt is very long due to extra content and the language model does not interpret exactly what you want. You probably saw this in the previous example. Then split the content with triple quotes at the beginning and end of the text from the rest of the prompt.
Example prompt: Can you summarize this article for me? “”” (paste your article text). “”” Start your answer with: The main points from this article are:
This method is called “content splitting” This example again uses an article from the internet: https://www.leraar24.nl/2619420/onderwijs-op-afstand-wat-hebben-we-geleerd/
Below is the language model's answer to the prompt with split content. We use triple quotes or three asterisks to separate the context of the question.
The main points from this article are:
- During the corona crisis, teachers had to devise strategies and techniques to make distance education as effective as possible, while parents took on a larger role in supporting their children.
- Important factors for successful distance education were organizing a structured daily schedule, maintaining open communication between teacher and student, and making well-considered choices about which lessons could be done online or on paper.
- Challenges relate to finding ways to support vulnerable students and the extra pressure placed on parents, especially those also working from home.
- There was a rapid increase in digitalization in schools, with the addition of new digital tools and methods. The article mentions the example of teacher Mandy de Vries, who, despite her limited teaching experience, was able to help her colleagues with her digital skills.
- The article also emphasizes the importance of ongoing contact and attention for students, especially those in special education, and offers practical tips to achieve this, both for teachers and parents.
- Finally, the article hopes that parental involvement in their children's distance education will continue even after the pandemic.
Technique 5 - ask for additional information first
You can give instructions to first ask additional questions before generating the answer. This helps the language model refine and specify your request.
Example prompt: “I want to draft a project proposal for developing a new mobile app for internal communication. The app should include features for news updates, an employee directory, and direct messaging. Before you write the project proposal, can you ask me additional questions to improve the proposal?”
This method is called “ask before answer prompting”
Below is the language model's answer to the prompt asking for additional information
Of course! Here are some questions I can ask to improve the project proposal: What is the primary goal of the internal communication app? (e.g., increase engagement, improve information dissemination, facilitate collaboration) Who is the specific target audience within the company? (e.g., all employees, specific departments, field workers) Are there existing systems or platforms the app must integrate with? (e.g., intranet, HR system, Active Directory) What is the intended budget and desired timeline for development and implementation? Are there specific security or compliance requirements the app must meet? (e.g., GDPR, data retention policy) What key performance indicators (KPIs) will be used to measure the app's success? Are there any ideas about the design or branding of the app? Should it align with the corporate identity? Is there a preference for the development platform (iOS, Android, cross-platform)?
With this additional information, we can get a better picture of the scope and requirements and thus draft a stronger project proposal.
Technique 6 - emotional prompt
You can give instructions to adjust the tone of the text to underlying emotions or to evoke a specific emotional response. This is useful for customer service, internal communication, or marketing.
Example prompt: “I need to draft an email to a customer who is unhappy about a recent service outage. The customer wrote the following: “”” …. “”” (paste the customer's text) I want the email to be empathetic, acknowledge the customer's frustration, and offer a constructive solution.”
This method is called “emotional prompting” This example uses a fictional complaint.
The full prompt is:
I need to draft an email to a customer who is unhappy about a recent service outage. The customer wrote the following: """Dear Sir/Madam,
I am very disappointed about the unexpected service outage yesterday afternoon. This severely disrupted our operations and caused productivity loss. We rely on your services and such incidents are unacceptable. I expect a clear explanation and appropriate compensation."""
I want the email to be empathetic, acknowledge the customer's frustration, and offer a constructive solution.
Below is the answer we received from the language model
Dear [Customer Name],
Thank you very much for your message and our sincere apologies for the service outage you experienced yesterday afternoon. I fully understand your frustration and the impact this has had on your operations. We deeply regret that we did not meet your expectations and our service standards.
The outage was caused by [brief, clear explanation of cause, e.g., an unforeseen technical failure in our data center]. We took immediate action and our technicians worked hard to restore the service as quickly as possible. The service has been fully operational since [time].
We take this incident very seriously and are already evaluating the cause to prevent recurrence in the future.
To somewhat compensate for the inconvenience, we would like to offer you [concrete proposal, e.g., a discount on your next invoice / a free upgrade for one month / a service credit]. One of our account managers will contact you within 24 hours to discuss this personally and answer any further questions.
Again, our apologies for the inconvenience. We value your business and strive to provide you with the reliable service you expect from us in the future.
Kind regards,
[Your Name/Company Name]
Technique 7 - requesting multiple versions
Here you ask for multiple versions of an answer so you have options. For example, you want the model to help generate different marketing slogans for a new product. There are many different angles possible, so you want several options. This can be effectively combined with setting a specific role for the model and a system prompt to further guide the model.
You can read more about roles in the relevant article. This example with multiple versions is at the very bottom of that article. It is not fully repeated here.
Technique 8 - combining with a system prompt
With a system prompt, you can give instructions to the AI model. It is an instruction you provide to the model to specifically shape the task or behavior of the model. You can specify what kind of output you expect from the AI model. It is a way to steer the model and prevent irrelevant or unwanted answers.
Some examples of simple and general system prompts:
"I want you to be a friendly and helpful assistant."
"Always remain professional and respectful in your language."
"Focus on giving concrete, to-the-point answers without unnecessary details."
The system prompt essentially sets the tone and expectations for the specific conversation or task you want to perform with the AI model.
You can also devise more advanced system prompts for specific tasks, such as summarizing a text: "Carefully analyze the given text and provide a concise but informative summary of the main points. Identify the core message and key details."
System prompts do not have the same influence on the outcome for every AI model. With chatGPT it works very strongly, with Claude also, with other models the influence of a system prompt is less. It is interesting to try and see which model works best for your input.
You can set a general system prompt in preferences, but it is best to devise these system prompts per task you want to automate. We call this roles.
You can read more about roles in the relevant article.
Annotating a passage
Select text in a completed chat message and choose Annotate. Write a question or modification instruction in the window, for example ‘Make only this sentence simpler’, and choose Save. You can annotate multiple passages this way. Then send; an extra prompt in the input field is not necessary but allowed. The passage remains saved with the comment, even if you switch to another window in between. You can remove an annotation before sending. On mobile, press and hold the text to select.