ChatGPT At Skyeng: How We Achieve Our KPIs And Work More Efficiently


We share our experience working with AI, a guide for compiling prompts and using ChatGPT in a product.
Many of you have already used neural networks – chatting with chatbots, generating pictures from a description, or converting an audio recording into text. But few people have begun using AI as a tool for their business, not just a toy, because even a simple, user-friendly chatbot is a powerful base you need to learn how to work with.

This article discusses how we turn neural networks from sprint to sprint into our assistants – fast and smart. We also share some of the materials of the closed workshop, which our CEO, Georgy Solovyov, conducted for Skyeng corporate clients.

If the customer fills out the brief crookedly, the performer is unlikely to solve the problem because, without a clear technical specification, you know the result. It’s the same with neural networks: the ability to create prompts for AI is a whole science.

Even with the advent of the first ChatGPT, we at Skyeng took up the tradition of testing all neural networks, looking for ways to use them. At first, they did not help to solve real problems and seemed “stupid.” We gotta appreciate the power of AI when ChatGPT 3.5 was released last fall – it created the queries you will see next. And the fourth-generation neural network passed our internal tests and began to integrate into business processes.

Work On The Quality Of Requests

ChatGPT can look at the problem from different angles, as in the method of six thinking hats – it will take on the role of critic, analyst, or creator. To arrange a “council,” you can ask ChatGPT to provide different points of view and opinions of real people:

We communicate with AI in both Russian and English. The second method has a more profitable tokenization algorithm, breaking the text into components by a neural network. ChatGPT reads English requests by words and requests in Russian – by letters.

Russian pangram contains 53 characters, and English – has 43. The difference is not critical. However, the Open AI tokenizer needs 9 tokens to process an English phrase and as many as 70 tokens for a Russian one. Ouch!

To get a quality result, we use several life hacks:
1. We ask the neural network to write a prompt to determine what data it needs.
2. We structure the request: we separate its parts with punctuation marks so they cannot be confused. We use simple and understandable words and follow the rules of grammar.
3. Set the context – so the neural network will have more data to select the appropriate style.

Product Integrations

Even before the hype around neural networks, developers began to use ready-made libraries – with their help, they wrote from 50 to 90% (!). Neural network assistants – for example, Copilot, who writes code based on text descriptions – are a new stage in this process’s evolution. It remains for the programmer to check the operation of the neural network and make changes. This frees up time for more important and complex tasks.

The neural network can write, check the code, and comment on its edits. Let’s “feed” Chat GPT 3.5 a piece of the script with errors and ask for feedback:

The plans are to teach how to recognize the difference between British and American English, advise how to improve the wording, make a separate browser extension and add anti-plagiarism.

Sky Grammar is not yet available for students, but they have “Kesha,” – a virtual interlocutor based on GPT-4. He maintains an oral or written dialogue, analyzes the answers, and identifies errors. Kesha cannot (and should not!) replace a teacher, but this is a good help for self-practice.

Solutions For Internal Processes

In the same way, neural networks will not replace editors and designers for us. Many tasks require critical thinking, creativity, a sense of humor, and empathy—all of which AI lacks. But the neural network effectively performs typical tasks. This allows you to save a resource of specialists for tasks with higher value.

With the help of AI, it is easy to create standard texts – contracts and vacancies. But ChatGPT 4 is capable of more. Our editorial team experimented with it during Q2 and developed the Promt bank with Skyeng’s editorial policy in mind. We now write some letters and posts for social networks based on the ideas proposed by the neural network – they have to be significantly improved, but this is much faster than writing from scratch.

The text-to-image model of Stable Diffusion aids illustrators and designers. We trained it on our database of illustrations, and now we are trying to generate pictures in Skyeng corporate style from a text description and a similar picture. Theoretically, it will save time in creating pictures for exercises and banners.

TLDR: How To Get Started With AI

Too long, didn’t read is an English acronym in comments to lengthy texts. Usually, it is followed by a summary of what was written, compressed to one or two sentences. And now “TLDR: (link)” is also a ChatGPT command, allowing you to solve a similar problem.

Clearly state why you need the help of a neural network. Be sure to include the scope and structure of your answer and the purpose and limitations. The more specific the request is, the more useful the answer you will get.

Test different prompts, and ask clarifying questions. This will help get valuable information not in the original answer.

Also Read: The Role Of Chatbots In Document Verification And AI

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *