Examining Your Work Before Setting a Prompt Goal

A prompt goal, or the desired outcome you want to achieve with AI, often means 'replacing your original work with AI.' However, there's room to reconsider the concept of 'replacement.' Let's think of AI as an assistant working alongside you. If you delegate part of your work to an assistant, that's just 'replacement.' But you could also ask them to review your output from a different perspective. This isn't exactly 'replacement,' but it can definitely enhance the quality of your deliverables.

I recently found a great expression in the EO YouTube video above that perfectly explains this perspective. The difference between groups that use AI well and those that don't is that they "treat AI as a colleague rather than just a tool." While tools are used unilaterally by people, colleagues can communicate bidirectionally and exchange questions. It's not wrong to view AI as a high-performance tool that can perform your tasks. However, this isn't the way to utilize 100% of AI's capabilities.

Therefore, we don't need to confine our thinking to the framework of 'replacement.' We can examine all types and characteristics of our current 'work' and apply AI where it can be useful. I work as a Product Manager in an IT company, and this article is heavily limited to my personal experience. However, I believe there are commonalities across various roles in terms of research, data analysis, and logical storytelling. Given these job characteristics, I rarely use AI-generated images/videos. Instead, I'll focus on natural language comprehension/writing abilities, mathematical operations, and programming capabilities.

Based on my daily work, tasks can be broadly categorized into three types:

Exploratory Tasks (e.g., Internet Search)

• Definition: Activities like online research where only a broad goal is set and related materials or data need to be found
• Examples: Investigating potential market size for a new product, learning about unfamiliar work methodologies
• Characteristics: Search efficiency is determined by how well you find the right 'keywords.' Since you're researching an unfamiliar field, finding the right 'keywords' can be challenging

Analytical Tasks (e.g., Statistical Calculations)

• Definition: Analyzing given materials using predetermined methodologies or techniques
• Examples: Deriving similar user clusters from user behavior data patterns, classifying VOC data by relevant features
• Characteristics: The 'correct answer' for individual data analysis steps is predetermined. For example, when creating a statistical report from given quantitative data, the method of calculating statistics has a predetermined answer

Creative Tasks (e.g., Writing)

• Definition: Creating specific outputs by utilizing given materials, ideas from your mind, and work capabilities
• Examples: Writing project proposals, drafting sales email content
• Characteristics: Since there's no predetermined 'correct answer,' it's important to generate diverse and effective ideas

Varying AI Effectiveness by Task Characteristics

Summary of AI tools by task characteristics

Exploratory Tasks: Shortening Progressive Search Steps with AI

Exploratory tasks here refer to the process of finding answers by searching for materials when only a broad goal is set. For example, after launching a new product in the market, you need to evaluate its performance after some time. The criteria for performance evaluation vary depending on various factors such as your company's target market and product characteristics. Therefore, you can't find the answer you want with just the keyword 'performance evaluation criteria,' and you need to narrow down search results according to your situation.

Now that I'm beyond the junior level, I'm somewhat familiar with industry terms and methodology terminology, making research easier. However, just a few years ago, I didn't know these terms well, and finding appropriate search keywords wasn't easy. The less familiar you are with a field, the more you have to start with broad exploration. You also need to consider the time it takes to understand the exploration results and narrow down the direction for subsequent research. However, with AI, you can quickly progress through these intermediate steps at an impossible speed for humans and immediately look at the final results.

Recommended Tool (Simple Problems): Perplexity (Price: Free)

For relatively simple research, you can use Perplexity. Installing the Perplexity desktop app allows you to quickly launch Perplexity with a shortcut key. It's a shortcut that can be executed anytime regardless of the currently open program, so whenever you want to search something while switching between various work tools, you can immediately input your question into Perplexity.

Perplexity input screen that can be launched immediately with a shortcut key on my computer

Before AI, it was important to input search keywords well to get good search results. For example, keywords that must be included in search results are enclosed in quotes. Since it's fundamentally about 'finding' materials that exist on the web, solving the problem was always my responsibility to read and utilize the materials.

On the other hand, with AI, you just need to tell it your 'problem' rather than the materials you're looking for. It then comprehends your intention, searches through internet materials, and derives a solution. For example, let's say you're researching which Excel function to use to get your desired analysis results. You can find various Excel function usage methods like vlookup through Google search, but of course, the materials found through search aren't tailored to your data. With AI tools like Perplexity, you can immediately tell it your data content and analysis goals. It then not only finds relevant Excel usage methods on the internet but also tells you how to use them with your data. Since you get the solution immediately without intermediate steps, there's no need to use search engines. Moreover, even with Perplexity's free plan, simple searches are unlimited, making it sufficient to replace Google search.

Google search now shows AI-derived results first. Thanks to AI Overview, some analysis suggests it's become more difficult to expose your content in search results by building SEO.

Elon Musk recently posted on X that "AI will replace search engines," and I think anyone who has used Perplexity would agree. It has actually replaced it in my daily life. By the way, you can also easily access GPT with a shortcut key by installing the GPT desktop app, but Perplexity has the difference of always citing source materials. This is a great help in reducing AI hallucination problems, as you can directly check the original text when fact-checking is important. It also has the advantage of conducting research while checking the original text.

As shown in number 1 in the image, it provides internet links used as references.

Recommended Tool (Complex Problems): Gemini Deep Research (Price: 29,000 won)

For more complex problems, especially in the initial stages where it's difficult to grasp what to research first, you can use Gemini Deep Research. There's a big difference in the amount AI writes - with Perplexity or GPT, the maximum amount AI writes at once is about one A4 page. However, Deep Research finds all materials directly and indirectly related to your requirements and creates a report of about 10 pages. It adds internet material links as footnotes at the end of the report, utilizing at least 10 and up to about 30 materials. Simply put, it feels like it finds all materials related to this topic on the internet. Moreover, since it summarizes and explains the materials in report form using appropriate tables of contents and tables, you can complete hours of Googling with just one AI prompt. Nowadays, when I need to research something, I start by turning on Gemini Deep Research. Since you can get all available materials on the web at once, it's excellent for getting a sense of what you need to know.

The downside is that it's a paid service that requires a Google AI Pro subscription. However, since it includes Gemini Notebook LM support, I think the price of about 29,000 won is worth paying. I also use Gemini Deep Research and Notebook LM most frequently for work. For work purposes, you need tools that can read vast amounts of materials or handle complex problems, and among the tools I've personally used, Gemini Notebook LM is the only one that can do both.

Analytical Tasks: Reducing Bias and Time Consumption

The characteristic of 'analytical tasks' discussed in this article is that they involve 'repeating' actions with predetermined 'correct answers.' For example, let's say you're classifying customer VOC data by related features. Since VOCs continuously flow in, classification work must also be continuously repeated. The criteria applied to such classification work should be as consistent as possible. This ensures consistent decision-making. Business data analysis must be able to consistently apply objective criteria for each data analysis execution. And this is exactly where AI excels over humans.

Even in the VOC feature classification work mentioned earlier, human bias can easily be applied. First, there's the perspective unique to the person doing the classification, from how finely to divide the features to whether to classify them broadly. Simply put, it's common for situations like "Why was this classified this way?" to arise just because the person doing the classification changed. In this way, AI's advantages shine when processing large amounts of data with qualitative criteria that humans must judge. AI doesn't get biased or tired.

Quantitative Analysis is Better with Excel, AI is Suitable for 'Large-Scale Repetition' of Qualitative Analysis

One point to note here is that AI is suitable for qualitative analysis. For quantitative analysis, it's actually more appropriate to use statistical tools like Excel or Python. In my personal experience, when I gave AI statistics data full of numbers and asked it to extract insights, I often saw cases where detailed number calculations were wrong. For example, it concluded that "User 2's A data is the highest at 1,000," but when I looked at the original, either User 3 was first instead of User 2, or User 2 was first but the number was 900 instead of 1,000 - there were cases where specific numbers were wrong.

The principle by which generative AI creates outputs is a 'prediction' task of selecting the most probable text based on learned materials. In contrast, mathematical calculations are processes that reach correct answers through logical reasoning. Of course, generative AI performance is growing at a pace that's hard to keep up with, so this could change as early as next year. However, at the current point, minor numerical errors can occur, which is an unacceptable disadvantage for statistical analysis work. In fact, such quantitative analysis can be done quickly enough with Excel, so there's insufficient reason to use AI.

On the other hand, when qualitative analysis needs to be repeated on a large scale, the effectiveness of using AI increases dramatically. The 'classification' work mentioned earlier is exactly this type of work. For example, if you're an Instagram customer data analyst and receive a VOC saying "I want to put an emoji like a sticker at my desired location while watching a friend's story," how should you classify it? If you try to automate classification work with rule-based methods without AI, you would have to classify based on whether words appear in the VOC. The word "story" could mean Instagram stories, but it could also mean actual stories. Whether to classify it under the "story" feature or the "reaction" category could also vary depending on the classifier's thoughts.

In this way, work that required natural language comprehension of given data and needed human reasoning abilities for analysis criteria was previously work that had to be done entirely by humans. Now you can save time by having AI do the analysis work instead. Moreover, AI maintains its fast judgment speed without getting tired, so it's more effective when the data to be classified reaches hundreds or thousands of cases. It also has more diverse general knowledge and basic expertise than I do, so it might even produce better judgment results than I would.

Analysis Criteria Must Be Judgable at the Level of Widely Known Professional Knowledge

However, even within qualitative analysis, there are tasks that AI cannot replace. Analytical tasks are processes of (1) understanding given materials, (2) understanding material analysis criteria, and then processing (1) with (2). The fact that AI can understand (1) and (2) means it can understand at the level of knowledge widely known online.

This is where difficulties in utilizing AI for work purposes arise. Although generative AI has learned from vast amounts of data recently, there can still be work that requires knowledge AI doesn't know. For example:

1. 1Knowledge created within the last year
2. 2Industry-specific jargon or practices, etc., very narrow field expertise
3. 3Other knowledge that's difficult to explain in 'natural language'

When judging whether AI can handle this work, it's good to think of 3 as a criterion. To utilize AI, you need to write prompts, and being able to write prompts means you can explain the work in natural language text. Earlier, I mentioned that people who treat AI as a colleague rather than a tool can use AI better. In other words, if you can write work instructions to pass on to an actual human colleague, that work can also be entrusted to AI.

Recommended Tool (When Analysis Materials are Small): GPT, Claude, Gemini - Choose Based on Cost

So which AI tool should you use for analytical tasks? If the data to be analyzed is about 3 A4 pages, I think any AI tool is fine. Claude, GPT, Gemini, etc. can read 10 A4 pages of data. They can read it sufficiently. However, in my personal experience, even with just 10 A4 pages, I felt that AI couldn't 'read carefully.' So the insights derived often seemed somewhat vague and obvious, and I think about 3 A4 pages is the most conservative standard for maximizing the use of given data.

AI inference performance test results are always up and down, so there's not much difference between Claude, GPT, and Gemini. Therefore, I think you can choose the cheapest service considering AI service subscription costs. I use Gemini almost exclusively for work purposes because my company has introduced the Google enterprise plan, allowing me to use Gemini's paid services. Google AI Pro subscription includes Gemini Deep Research, Notebook LM, and using Gemini in Google Docs, so going all-in with Google is also a good method.

I particularly recommended these three services as examples because they can 'manage prompt instructions.' Claude Project, GPT Project, Gemini Gem fall into this category. The terms vary slightly by service - Project, Gem, Rule, etc. - but their roles are the same. It's a function to specify rules that AI should always keep in mind before starting an AI chat. It's good to use when you want to create a chat window optimized for specific purposes.

For example, you can upload "Our Company's UX Writing Guidelines.txt" to the prompt instructions and write rules like "Always check Korean standard spelling." The chat window with these instructions applied always uses the uploaded .txt file and Korean standard spelling. In this way, prompts that serve as fixed 'guidelines' can be uploaded to 'prompt instructions.' And for materials or data given for your current work that need to be analyzed each time, you can pass them as input data to the AI chat window to utilize AI more efficiently.

Recommended Tool (When Analysis Materials are Large): Gemini Notebook LM (Price: 29,000 won)

For large amounts of data, dozens of A4 pages, I recommend using Gemini Notebook LM. I've often seen 'too long conversation' errors when giving large files to Claude or Cursor, but in my personal experience, Notebook LM is the only one that can comfortably handle large data.

When you open Notebook LM, you first see the Source upload window. It's convenient for utilizing recent materials as it can read content well even with just Website or Youtube links. It also supports file uploads like PDF, txt, and Markdown.

Source upload window that appears first when opening Notebook LM

Unfortunately, you can't upload Google Sheets, only Google Docs and Slides are supported. For Google Sheet analysis, there's a method to use Gemini by opening the Gemini prompt window directly in the Google Sheet screen. Notebook LM doesn't support CSV uploads either, so I've had to convert them to .txt to upload, but when I tested the prompts, it seemed there was no major obstacle to AI understanding the content even as .txt.

You can activate the Gemini chat window by clicking the Gemini logo button in the top right of Google Sheet.

When you upload a Source to Notebook LM, you can see the main screen like the screenshot below.

Notebook LM main screen

1. 1You can check whether to use the uploaded Source. It's useful when you want to narrow down the materials that AI should utilize in this prompt, even if you've uploaded multiple Sources at once.
2. 2You can input prompts in the chat window in the center of the screen. You can check/uncheck the materials to utilize from the Source and input the prompt. Even if the Source is a very large file, you can get prompt results of decent quality.
3. 3The results generated by the prompt can be saved as Notes. Notes are managed in the Notes menu visible in the bottom right of the screen. You can also upload Notes as Sources again. In other words, you can narrow down analysis results for large Sources by repeating the process of analyzing Sources with AI → saving analysis results as Notes and uploading them as Sources again → analyzing Sources.

Creative Tasks: Filling in Aspects You Might Have Missed

While Analytical Tasks are Automation, Creative Tasks are Gap-Filling

The creative tasks I want to discuss in this article refer to the act of proposing new ideas or creating new outputs by logically connecting given materials. Among Product Manager tasks, writing proposals (one-pagers, PRDs) is a representative example. It's an act of creating something that didn't exist before, without predetermined correct answers. Therefore, how many diverse ideas you can sufficiently try is important.

When ideating, it's easy to think of very creative ideas, but even if it's not to that extent, there's room to get help in the ideation stage. When you need ideation, if you ask someone for help, they often give you their own ideas. At this time, it's rare for their ideas to be something I couldn't have thought of. Often, after hearing them, I think, "I could have thought of that, why didn't I?" Whether it's due to time constraints or focusing too much on the material at hand, it can be difficult to increase the number of ideas on your own.

The difference in utilizing AI between analytical and creative tasks lies in this 'gap-filling.' Analytical tasks are closer to automation because the goal is to complete predetermined tasks more, faster, and without bias. On the other hand, for creative tasks, it's important whether you've sufficiently reviewed all possibilities. As humans, it's easy to have a narrow perspective due to various variables, so borrowing someone else's perspective is the easiest solution. AI with human-level comprehension/reasoning abilities can play that other person's role.

Creative tasks can be divided into two main aspects:

1. 1Ideation: The initial stage where it's difficult to come up with ideas
2. 2Review/Evaluation: The final stage where gaps are found in your output

When it's difficult to come up with the first idea, when you need to increase the number of ideas, or when you want to save time by having AI do what you could do, you can make work more efficient by having AI do ideation. At this time, rather than asking for vague ideas, it's important to provide appropriate guidelines. Just as humans can be more creative when there are constraints on their thinking, AI's accuracy for individual ideas increases when ideating within a set framework. Also, AI prompting basically requires providing the format and appropriate examples of the desired output to get decent results.

On the other hand, after completing the previous stage (whether with AI's help or not), you can ask AI for opinions on your first draft. You can provide 'the definition of an ideal output' as a guideline in the prompt and ask it to review/evaluate your output based on this criterion. For example, you can find materials like 'the definition of a great proposal, Amazon 6 Pager writing method' online, make them into prompts, and have AI review your proposal.

Even if AI's review results are somewhat obvious, they can help with 'gap-filling' by finding points you might have missed. You might know in your head that you need to fill all items 1, 2, and 3, but end up spending too much time on item 2. This means items 1 and 3 unintentionally get less time and end up with weak content. When AI finds these gaps and suggests how to fill them, you can improve the quantity and quality of your work as if getting feedback from a colleague.

Recommended Tool (Review/Evaluation Tasks): Cursor (Price: $20)

The same rules as analytical tasks apply to creative tasks. If the total amount of materials to input to AI is dozens of pages, use Notebook LM, and if it's a small amount, use GPT, Claude, Gemini, etc. And it's fine as long as it's not work that uses very specific tacit knowledge. However, I want to recommend Cursor, an AI coding tool, for review/evaluation tasks.

As explained in another post, Cursor can be expanded to use as a 'prompt editor.' Forget the preconception that it's an AI coding tool, and think of it as having an AI chat window attached to your computer's file explorer. You can use it for purposes completely unrelated to coding. Cursor's special feature is that when you open any folder in your file explorer as an editor and give requirements to AI, AI directly modifies your files. Here, 'files' are usually source code and 'requirements given to AI' are requirements needed for programming development.

Cursor doesn't use AI models specialized only for coding. It supports GPT, Claude, and Gemini, and the key UX is that it immediately applies improvements to your source code from these models without you needing to edit directly. Therefore, it's entirely possible to apply 'requirements for general writing, not coding' to 'general writing files, not source code.' For example, you can give your 'blog post draft' and apply the prompt 'revise for readability.'

cursor screen structure

1. 1When you open a specific folder from your computer in Cursor, you can see the file list on the left side of the screen. It serves the same role as a file explorer.
2. 2When you click on a file you want to review with AI, you can see the content in the center of the screen. Here, you can specify the desired area with mouse drag and press the shortcut key (Command + i / Ctrl + i) to attach it to the AI chat window.
3. 3You can see the AI chat window on the right side of the Cursor screen. You can see the materials attached to this chat window to the right of the @ button. You can check the attached part (line 30 of post.md file) from step 2.

Asking cursor to modify my file

1. 1Tell the AI chat window what guidelines to use to review/evaluate your output.
2. 2Here, using Agent mode among (1) Agent (2) Ask (3) Manual modes allows AI to directly modify your files.

If you want to see only the prompt results without having AI touch your files, you can use Ask mode. Like other generative AI services, it only allows conversation with AI in the chat window without AI accessing your files.

Finally, Manual mode allows AI to directly edit your files just like Agent mode, but with the difference that it can only edit files you've directly specified in the chat window. Here, specified files refer to materials that can be added to the right of the @ button in the chat window. You can specify files or parts of files, and you can also add materials outside your folder like website links. Agent mode can 'automatically refer to other files in your folder as needed' without you specifically telling it which files to look at in the chat window. Here, folder refers to the path you've currently opened with Cursor. Agent can only automatically read files visible in the Cursor left list. Cursor is originally a programming tool, and since source code files are closely connected, this feature of AI automatically looking at other files is very useful.

As explained in this article, if you're only using it for writing purposes, Manual mode might be sufficient. However, Agent mode is the default and recommended setting for Cursor, and even if it's not source code, there can be many cases where multiple files are related to each other. Agent mode has the advantage of not needing you to specify this one by one, so it's more convenient to use Agent mode if possible. (It's not like you lose anything by using it!)

cursor directly modifying my file

1. 1You can see the prompt results, that is, AI's review opinions of your output, in the AI chat window on the right side of the screen.
2. 2In Agent mode, Cursor directly modifies your files, and you can choose whether to apply this by pressing the Accept button at the bottom of the screen. Pressing Reject returns to before the edit.
3. 3Instead of Accepting the entire file, if you want to judge partially, your original text is shown in red and AI's edited version in green. You can press Accept/Reject buttons for each part.

Conclusion

The AI services recommended in this article are all limited to my personal experience. There might be better services I don't know about, or new ones might be released. These days, AI news is pouring out at an unprecedented speed. That's why you've probably heard at least once that it's difficult to keep up.

Nevertheless, the situation where GPT, Claude, and Gemini have established themselves as major players in text-centric tasks is unlikely to be shaken in the near future. Google wasn't even a comparable alternative to GPT until recently. I remember hearing that its performance was excellent with the recent Gemini 2.5 model. And then the situation changed so much that LinkedIn and YouTube feeds were full of Google I/O talk. Unlike OpenAI for GPT and Anthropic for Claude, Google holds 'B2C, B2B services' like search engines, document tools, and email, so I think it only has faster growth ahead.

If the major AI models that show decent performance at any time are determined, the process of deciding what and how to do with those AI models becomes more important. Tools can change, but what you intended to do with these tools doesn't change. Which company's model is the best can change, but the fundamental principles of 'AI prompting' don't change. Therefore, you should focus on designing how to structure and link AI prompts to reach your goals. Once the structure is in place, you only need to swap out the tools or AI models used at the endpoints.

In the previous post, I defined prompt goals as the output of tasks to be performed with AI. And I explained that prompt workflow refers to how many times AI needs to be executed to reach that output and how to link each execution result. Particularly in analytical tasks, prompt workflow design is important.

For example, let's say you have 1,000 customer VOCs to classify. You could give all 1,000 to AI at once for classification, divide them into 100 each for classification, or even give them one by one for the most precise classification. As mentioned several times earlier, as the amount of materials given in one prompt increases, AI can't 'read carefully' and understands vaguely. So if you want to fully utilize the given data, it's better to give as little input data as possible in one prompting.

The problem is that when the number of prompt executions reaches hundreds or more, you spend more time pressing enter in the chat window for prompts than the time saved by prompts. To solve this, you need to be able to use generative AI as an API.

Earlier, I said that for analytical tasks with small amounts of materials, you can use any of GPT, Claude, or Gemini. In the summary image at the beginning of the post, I marked GPT as the recommended service among them. This is because among the 'generative AI services used as API' that I've tried, GPT API was the best. Analytical tasks are likely to become automation work, so being able to use generative AI as an API opens up much wider possibilities. You can make it automatically repeat prompts 100 times with one button click, so you can try more diverse approaches.

In the next post, I'll discuss how to use prompting multiple times and link them to further unleash the potential of your data. I'll organize the advantages and usage methods of using generative AI as an API from the perspective of non-developers like me.