Best practices for developing an effective GEN AI prompt
In this article, we offer you some of the best practices to make your GEN AI prompt effective when capturing your customers' intentions.
Developing an effective prompt in GEN AI can make the difference between obtaining accurate and clear results, or facing confusing and unhelpful responses. Here are some key tips to ensure your prompt is well-structured and optimized for best results:1. Information Quality:The foundation of any effective prompt is the quality of the information provided. Make sure the data is well written, free of misspellings, and clearly structured. Use subheadings and lists to organize information effectively.2. Organization:Organize your prompt into clear and concise sections. Divide the information into subheadings and lists that define the data to be trained. This will help the bot process the information more effectively.3. Appropriate Format:Use file formats that are compatible with the platform, such as plain text in PDF format or the one required by the platform. Avoid using tables or other types of documents that may make it difficult for the bot to read the data.4. Detailed Structure:Provide the bot with a clear context so that it understands how it should act. Divide the prompt into sections that include information such as special tasks, formats, limitations, etc. The main ones we recommend are:
Context: here you will describe how the bot should behave, the type of voice it should have, and what it should report. Example: Act as a car sales expert for company "x", and answer customer inquiries in a friendly and helpful manner. If someone asks for information that does not correspond to the company, simply answer that I do not have access to such information.
Format: here you will describe how the bot should respond in terms of the format of its responses. For example: Respond in a message of no more than 3000 words, without emojis, and organize the information in lists and paragraphs of no more than 3 lines.
Restrictions: here you will highlight everything that you do not want the bot to answer. For example: If someone asks about competitor products, simply answer that you cannot give that information. You cannot be creative with the answers.
This will help the bot better understand its functions and process information more effectively.6. Relevant Q&A:Consider the questions that are frequently asked by your customers and those that should be answered in a certain way. Train the bot by including these questions and answers in the 'Questions and Answers' section. This will help reduce incoherent responses and improve the user experience.7. Image Handling:If you need to include images, upload them to the cloud and generate a link with the appropriate extension (.png, jpg). Then, include these links in the questions and answers section so that the bot can access them when needed. Documents are uploaded in the 'files' section and are uploaded from the computer, these must not be uploaded to the internet.By following these tips, you can develop an effective GEN AI prompt that maximizes the accuracy and usefulness of your results.
🧠 Consistency validation in knowledge bases for bot responses
When a bot uses astructured knowledge baseto respond to queries such as prices, descriptions, or product availability, it is essential to ensure thatall fields are correctly updated.
A small mismatch (for example, updating only the description but not the price) can cause erroneous responses in production.
✅ Checklist for publishing updates to the Knowledge Base
Before saving or publishing changes, review:
Did you updateall relevant columns?(📝 Ex.:name,description,price,stock,URL, etc.)
Did you updateall relevant columns?(📝 Ex.:name,description,price,stock,URL, etc.)
Were the data modified in thecorrect version of the bot's month?
Were the data modified in thecorrect version of the bot's month?
Was a review of the complete dataset (rows and columns) done before saving?
Was a review of the complete dataset (rows and columns) done before saving?
Did you test the behavior in the bot by simulating the intention or keyword?
Did you test the behavior in the bot by simulating the intention or keyword?
⚠️ Common mistakes you can avoid
| Frequent error |
| Why does it occur? |
| How to avoid it |
| Only the product description is updated |
| A cell is edited in the row without reviewing other related columns |
| Use filters to view the complete content by product before saving |
| The correct document is edited, but not published |
| The knowledge base remains as a draft or is not linked to the correct bot |
| Confirm in the bot's menu that the active Knowledge Base is the appropriate one |
| Duplication without cleaning |
| A dataset is duplicated without deleting previous values |
| Review and clean the columns before duplicating or copying |
🧪 Real example
A customer reported that the price of "article1" was incorrect. The description had been updated, butthe price columnmaintained the previous value.
👉 Solution: the price field in the March database was updated and verified in the flow. Result: the bot now responds correctly with $500000.
🧩 Final recommendation
If you work with multiple monthly or campaign versions of your datasets, implement aclear nomenclatureand a control table that indicates:
Date of last edit
Date of last edit
Published version
Published version
Responsible user
Responsible user
Which fields were modified
Which fields were modified