Hello everyone, my name is Roman Kagan and I have been a CTO in fintech for the last 5 years, leading a team of 30 people. Programmers have long and firmly used AI in development: in writing code, documentation, and tests. This saves time (and money), and allows us to quickly develop and release products even with a small team. Some programmers even use AI to hold two full-time jobs because they have (surprise!) a good understanding of how to use them. When I mentioned this in a conversation with a fellow entrepreneur, I was surprised that almost no one outside the programming community either does not use AI, or does it incorrectly I decided to fix this and try teaching a friend of mine, a psychologist who runs his own corporate consulting company, how to use AI in his work. I inquired about their activities and realized that the simplest way I could help was through content marketing. My task was not just to address a specific issue but also to teach my friend how to use ChatGPT effectively, including how to craft quality prompts for various tasks. We had a call and worked together for two hours. As a result, my friend fired his copywriter and began saving as much as $1200 a month on content writing. # What were the doubts about the plan to replace a live specialist with AI? My friend liked the tone of voice of his copywriter, everything was calibrated and the process was well-established. Therefore, he doubted the success, but agreed to allocate 2 hours. # In what cases is AI definitely unable to replace a human? What we call *creative work* is divided into two very different categories: **art** and **craft**. My assumption was that a good copywriter's ability isn't just about creativity, but rather skills honed through experience. I proceeded from the fact that the skill of a good copywriter is not made up of his creativity, but rather his skills honed through experience. And this is: * The ability to write briefly and clearly * Analyzing the target audience of readers * Explaining complex things in simple language * Effectively structuring complex subjects for easier comprehension * Quickly researching a topic to distinguish valuable information from the mediocre **Literature** **demands great creativity** and constant search for new expressive means, as writers aim to evoke complex emotions and craft entire worlds — where talent, not just experience, is crucial. No AI can replace a human in this truly creative work. Conversely, corporate blogging **is a craft** and requires the skill to simplify complex topics, honed through regular practice. AI **is wonderful in that it learns very quickly**, trains and can analyze huge amounts of data in seconds. Therefore, if you load into AI all the experience that a copywriter needs to live through to become a good author for business, AI is quite capable of becoming one. And in a very short time. And therefore our task is to train AI and turn it into our employee. # How My Friend Worked with a Live Copywriter My friend regularly publishes case studies from his practice. First, he dictates notes into a recorder, then sends this to a copywriter. The copywriter translates the voice into text, structures it, and writes an article based on this. Then he adapts it for different social networks. The pricing math is straightforward. It takes about 30 minutes to transcribe one material, 4 hours to write and edit, 2 hours to adapt for platforms, layout and posting. 3 materials are produced in a week. In a month, my friend spends about $1500 (4w * 3 articles * 6h * $20/h). This is how we aimed to reduce this expense. # Don't Use OpenAI's Chat Interface: What is Playground There is a basic life hack if you don't want to get a poor result. Use ChatGPT-4, it is available by subscription. And I recommend not using the chat interface available to regular users at the link [chat.openai.com](https://chat.openai.com/). Instead, use [Playground](https://platform.openai.com/), which is available for developers. This is important for a good result because: * The latest gpt-4-turbo-preview model is available there, which has a larger context and works better with it * There is a temperature parameter that affects the "creativity" of chatGPT * There is a more convenient interface for working with prompting, you can specify a system prompt that will trigger for each request, and there is also the possibility to save presets. ![img](88iylneduuqc1 "This is how Playground looks like") >⚠️ Payment in Playground is made for each request. You may need to top up your account. For experiments, $5-10 should be enough. Now to the writing of the prompt itself. # How not to write a prompt How most people understand prompting: >Act as a professional corporate psychologist write an article using notes In this scenario, you'll end up with the he kind of bullshit that'll make you cringe: it's blatantly obvious the post was penned by a neural network. # How to Properly Write Prompts How I see prompting: ![img](0o8g8mrhuuqc1 "A schema on how to break down a task for AI. The prompts themselves are marked with ?, they are simplified for the sake of schematics") A skilled copywriter divides the task into distinct steps: analyzing the audience, handling data, outlining the article, and so on. To make ChatGPT a proficient copywriter, it too must break down the task into specific steps. Namely: * have it focus on each stage independently * create a chain of requests, where the result of one prompt is fed into the next Our task is to create in this way one giant, detailed prompt, according to which AI will write us an article. # How we broke down the task of writing an article for ChatGPT It was important for us to preserve the Tone of Voice of the current author. I have outlined the following steps for the chat: * Work out the author's Tone of Voice * Develop the author's Persona * Work out the reader audience * Work with notes from a psychologist * Gathering all the information together in the final prompt The final prompt looks quite massive, more than 20 thousand characters. This was far from my friend's idea of how to write prompts for ChatGPT. But the real key to using AI effectively lies in accurate and detailed prompting. # Step-by-Step Guide to Training ChatGPT to Become Your Copywriter **Step 1: Sampling** Before anything, we gather between 6 to 10 existing posts from the Author (my friend's copywriter). It’s crucial to choose a diverse set of topics to avoid topic-specific bias in our analysis. This data will be used to understand and mimic the user’s unique tone and style. Ideally, after collection, split these materials into two equally sized datasets (simply choose the good ones, split into two batches and keep in Notion for further use) One dataset will be used to extract the Tone of Voice. Another dataset will be used as examples (few-shot prompting) to improve accuracy. **Step 2 Building Tone of Voice** The process begins with the need to evaluate the author's tone of voice accurately. Even the styles of Leo Tolstoy and Hemingway can be broken down into formal criteria. To mimic the tone of a copywriter you admire, first compile these criteria — something ChatGPT is capable of doing. So, we're submitting a request to the model to "Build a framework to evaluate voice tone." This action generates a comprehensive set of criteria that can guide the evaluation of an author's tone. If necessary, further requests can be made to refine or expand this framework, ensuring it aligns closely with the project's specific needs. >? After executing the prompt above, you can ask the model several times, "Do you have anything to add?" This allows you to pick up more items for your framework. After ChatGPT generates the criteria, keep them in notes. **Step 3. Applying the Tone Framework** The next step is to help the AI understand the tone of voice of the initial author in order to learn how to impersonate them. So, we combine the formal criteria from step 2 and the examples of the author's best articles in the next prompt. **Prompt to Describe Author's Tone of Voice** >**<<Context>>** >I’m automating blogpost writing using LLM for telegram platform. I have samples of telegram posts, information about the author, target audience and transcription of voice notes containing idea and content of the future post. >The plan is: >Build the Tone of Voice based on samples provided >Build Persona profile for the Author >Build Persona profile for the Audience >Parse voice notes into the text >Combine all together in the final post. >**Examples:** [put 3-5 posts here] ><<**Objective**>> >we are now executing #1 from the plan. Your goal as a writing and literature and media expert is to describe the overall author's tone of voice. Use the Tone of Voice Framework and provided examples of blogposts. Do NOT use references to specific posts, I need the overall overview. ><<**Style**>> >Act as a professional copywriter and literature expert ><<**Tone**>> >Concise and straightforward. Every word is a precious resource. >**<<Audience>>** >Firstly you response will used by another LLM, but before it will be verified by a IT professional. >**<<Responce>** >Tone of Voice Framework [put the result you got during the previous step] By analyzing these examples through the lens of the established criteria, it's possible to populate the framework with concrete data unique to the author's style. This translates the theoretical aspects of the tone framework into a practical, data-informed profile of the author's voice. This profile is a key artifact used in the subsequent steps of creating posts. >? For a large framework, consider analyzing each section individually or in groups of 2-3 sections, instead of having ChatGPT complete all fields in one request. **Step 4: Building Persona Framework** Creating personas is key to making content that fits both the writer and the audience. This means building profiles for two personas: one for the writer and one for the audience. Making the writer's profile is just as important as the audience's. For instance, ChatGPT can tell if posts are by men or women because it's learned from millions of internet articles. Sometimes we don't spot the difference, but AI does. Creating a persona framework helps make the content more personalized and human-like. ChatGPT can help build this framework, which needs details like age, education, income, hobbies, job stuff, and so on. With this framework, we can create detailed personas using data like public social media stuff and personal interactions. I went with the framework I already had. **Prompt to create Persona** ><<Context>> >I’m automating blogpost writing using LLM for telegram platform. I have samples of telegram posts, information about the author, target audience and transcription of voice notes containing idea and content of the future post. >The plan is: >Build the Tone of Voice based on samples provided >Build Persona profile for the Author >Build Persona profile for the Audience >Parse voice notes into the text >Combine all together in the final post. >Persona information: [Raw information about the Persona from social media and other resources] ><<Objective>> >we are now executing #2 from the plan. >Create a persona profile that will be used to write blog posts on behalf of a real person. This persona will be utilized in a Language Learning Model (LLM) to simulate human behavior and perform various tasks, including writing. ><<Style>> >Act as a professional product manager and marketing specialist ><<Tone>> >Concise and straightforward. Every word is a precious resource. ><<Audience>> >The output will be used by LLM as an input and context for another task. It will be also reviewed by human product manager before sending to LLM. ><<Responce>> >Use the following format to form the answer. ><aside> ❓ If you don’t have information about the Persona to fill certain feature try to imagine something yourself but then put <review> tag at the beginning ></aside> >Persona Template >Name: [Persona Name] >Age: [Age] >Gender: [Gender] >Location: [City, Country] >Education Level: [Highest Degree Obtained] >Income Level: [Annual Income or Range] >Professional Background >Industry: [Industry they work in] >Job Role: [Current Job Title] >Years of Experience: [Years in the industry/job] >Psychographics >Personality Traits: [Key personality traits] >Values and Beliefs: [Core values and beliefs] >Interests and Hobbies: [Main interests and hobbies] >Technological Competency >Familiarity with Technology: [Level of comfort and familiarity with technology] >Digital Skills: [Level of digital literacy] >Adaptability to New Technologies: [How quickly they adapt to changes or updates in technology] >Goals and Motivations >Primary Goals: [Main goals they aim to achieve with the product] >Secondary Goals: [Other goals that are not the main focus but still relevant] >Challenges and Pain Points: [Specific challenges they face that the product could solve] >Behavioral Patterns >Usage Scenarios: [Specific situations or contexts in which they would use the product] >Purchasing Behavior: [How they make purchasing decisions] >Media Consumption: [Preferred media channels and content types] >Expectations and Preferences >Product Expectations: [What they expect from the product in terms of features, benefits, etc.] >Brand Preferences: [Preferred brands and why] **Step 5: Making the Article Comprehensive** Now comes the crucial bit — analyzing the professional knowledge and experience my friend usually expresses through his social media posts. Without these, any article turns into well-written bullshit, whether it's ompiled by a human copywriter or an AI (like all that pointless stuff people just slap together from Google). My friend used to send voice or text notes to his copywriter and answer clarifying questions until everything was clear. This made the article thorough and engaging, covering the topic comprehensively. We can use any voice-to-text tool (Google Meet or Zoom) to get the raw transcript. Now we can improve it by emulating the audience persona behavior to predict, ask questions, and express concerns. So, we used ChatGPT to formulate relevant questions that an audience member might have after going through the content. My friend responded to these in written form or via additional voice notes. **Step 6: Combining Elements for Post Creation**: Once all elements: * Tone of Voice, * Author's Persona, * Audience Persona, * and Enriched Content are ready, the next step is to combine them to create **the prompt**. This is done by sending a request to ChatGPT that merges the tone, personas, and content into unified posts. We improve the model's precision in replicating the author's style by including examples from a second dataset of posts (not used for tone analysis) through few-shot prompting. This gives more context to ChatGPT. **Scheme of the Final Prompt** >**<<Context>>** >I’m automating blogpost writing using LLM for telegram platform. I have samples of telegram posts, information about the author, target audience and transcription of voice notes containing idea and content of the future post. >The plan is: >Build the Tone of Voice based on samples provided >Build Persona profile for the Author >Build Persona profile for the Audience >Parse voice notes into the text >Combine all together in the final post. >**<<Raw transcript>>** ><<**Objective**>> >We are now executing #5 from the plan. >Act as PERSONA write a blogpost following TONE OF VOICE for AUDIENCE and using the information provided in RAW TRANSCRIPT. ><<**Style**>> >Act as Author >Author**:** [Structured information about Author’s Persona] ><<**Tone**>> >[Structured information about Author’s Tone of Voice] >**<<Audience>>** >[Structured information about Audience Persona] >**<<Responce>>** >[Examples] # Results We generated a new long-read for the Telegram channel, and my friend posted it with a few minor edits. Frankly speaking, it received more positive feedback from the subscribers than the previous ones written by humans, and we got zero comments suspecting the article was AI-generated. My friend practically stopped paying the copywriter. He himself generates several versions of draft posts, selects the best one, and sends it to a person for proofreading, layout, and publication, which all in all now takes 1 hour of time. And saves about $1250 a month (1500(1 -*1/6)). Feel free to take the tips from the post and apply them for yourself. If you encounter any issues, just drop a comment here, and I'll offer advice and make any necessary corrections. # Outro If there are tasks in your business that you'd rather not handle or pay for, feel free to let me know in the comments. If the task is straightforward, I'll gladly advise you on how to delegate it to AI like a pro. And I provide personal consulting - my email ([roman@nirvana-privacy.com](mailto:roman@nirvana-privacy.com)) or Telegram (@romankaganov)