Working with AI: Boosting Productivity#

Now that we have covered the concepts and the tools, let’s talk about how they can change your own work right away. This chapter is about changing how you work alone before you build a product alone. In a one-person startup, the founder’s time is the most expensive resource, and where that time goes decides the outcome.

As we saw in the previous chapters, services like ChatGPT, Gemini, and Claude can be used for information search, summarization, translation, writing, proofreading, code writing and review, data analysis, image generation, idea generation, and more. Each service has its own partial strengths and weaknesses, but none of them are absolute and they shift over time, so it is best to try several and pick whichever fits the problem at hand. The first half of this chapter walks through those uses one by one; the second half goes beyond the chat window and covers handing an entire workspace to an agent.

What you can do in the chat window#

Information search#

Traditional search services take keywords from the user, list related data, and leave it to the user to browse the results and find what they need. With an AI service, you ask in natural language, the way you would ask a friend, and get back a single well-organized answer. Even when you have no idea what to search for, you can simply describe the situation: the AI makes sense of it, tells you what to do, and asks follow-up questions when needed in order to respond more precisely. That said, AI has a hallucination problem and its answers sound quite convincing, so when you need accurate information, do not take the response on faith; verify it.

Say a new hire at an e-commerce company is told to “add information to the order spreadsheet so it is easy to see what product each order is for.” If they do not know that a separate product-information sheet exists, and they are not familiar with Excel functions, they will not even know what to search for. But if they describe the situation to an AI, “I was told to add information to a sheet that records customer ID, shipping address, product code, and quantity, so that it is easy to tell what the product is. How should I do this?”, they get a step-by-step walkthrough: prepare a product-information sheet and look up product names with VLOOKUP. Then they can learn what VLOOKUP is from the official documentation. Finding out “what to search for” is the biggest value of AI search.

Summarization#

We live in an age of information overload. Some of that information deserves careful reading, but for much of it, even after reading a long article, only a sentence or so sticks. In that case it is more efficient to read a summary first and dig into the details only when needed. You can paste the URL of a web page or attach a document file in a chat-style AI service and ask for a summary, or ask an AI browser to summarize the page you are about to read. I run the AI news briefs on this blog by having an agent read multiple sources every week and draft a summary, which is “summarization” turned into automation.

Translation#

Translation was in widespread use long before ChatGPT became a sensation. Two things have changed. The first is context and tone. Unlike a dedicated translator, instructions like “in a business email tone” or “keep technical terms in the original language” actually work. The second is multimodality. Speech recognition, translation, and speech synthesis used to be separate models, which made the whole thing slow and lowered quality. Now that a single model can listen to speech and reply directly in another language, something close to real-time interpretation is possible. The English edition of this book was produced the same way: an agent translated it and I reviewed it.

Writing and proofreading#

Writing is one of the most powerful abilities of LLMs. Novels, screenplays, reports, proposals, naming, planning ideas: they understand the user’s request and context and write accordingly. But if you already have the exact text in your head, AI may not be the right fit. To get a specific result, you have to write an equally specific prompt, and that can take longer than writing the thing yourself. So it makes sense to use AI when you have only the outline and are fine leaving the details to it, when you want ideas from what the AI writes, or when you need a first draft of a document where format matters, such as a business email or a report. It is also useful for checking the grammar and flow of what you wrote yourself.

Code writing and review#

AI’s ability to write code works on the same principle as its ability to write prose. It can write because it was trained on enormous amounts of text, and it can code because it was trained on enormous amounts of code. The difference is that code is more structured than prose, so there is relatively less room to wander off, and you can verify correctness with tests. I believe developers are the group that has benefited the most from LLM-based AI services.

The term “vibe coding” became popular. The concept is that you can build an application just by describing what you want in natural language. There is still a gap between the ideal and reality, but it is certain that the time spent “typing code by hand” has dropped dramatically. For people unfamiliar with programming, though, the wall is still fairly high. Getting a first draft of the code from the initial request goes smoothly, but when that draft differs from what you want, steering it in the right direction is harder than it looks without a basic understanding of programming.

AI is also very good at explaining code. Taking over someone else’s code and having to update it was one of the situations developers hated most, and by the time a question comes up after the handover period, the previous owner is often gone. AI, on the other hand, will read and explain the code any time. See the coding agent chapter in Part 3 for how to use it in earnest.

Data analysis and visualization#

Professional BI tools like Tableau and Power BI are powerful, but they have a steep learning curve and cost money. With a chat-style AI service, you can attach an Excel file or a Google Sheet and ask, “For each of ‘in-progress tasks’ and ‘planned tasks’, sum the ‘development period’ by ‘category’ and draw a bar chart,” and you get the chart and an explanation without learning anything new. It falls short of professional tools in precision and reproducibility, but for an early-stage one-person startup, the realistic order is “look quickly with AI first, and bring in a professional tool when the analysis needs get complex.”

Inside Excel, Microsoft Copilot handles tasks like creating charts, writing formulas, and building pivot tables from natural language. The difference from a chat service is that you do not have to explain the current situation (context) separately. With a chat service you have to upload the file or describe the table before asking; Copilot is integrated into Excel, so it automatically understands the open table and edits it on the spot. This is what “the product supplies the context for you” in Part 3 means.

Image / video generation#

You can produce draft-level visual content such as marketing banners, product images, app icons, and short social videos without a designer. It is also common to chain the two: feed a product image made with an image-generation AI into a video-generation AI to create a promotional clip. All of the illustrations in this book and on this blog are produced with image models. The limits are as described in Part 3: they capture a mood well, but precise diagrams are hard, and the realistic workflow is to generate several candidates and pick one rather than expecting the right result on the first try.

HTML page generation#

With AI you can quickly build HTML pages for all kinds of purposes. “HTML page” here does not mean “formally developing a website”; it means a simple screen that runs directly in the browser (input forms, buttons, a results area, and so on). It is especially useful for validating an idea before committing to costly, time-consuming real development, or for building a small tool that solves an immediate annoyance.

The first use is prototyping to communicate product requirements. A written spec can lay out requirements as text, but it struggles to convey the “execution flow” vividly: which button the user presses and which screen they land on. Hand the spec to an AI and ask it to “build a prototype HTML that shows this flow,” and you can quickly produce and share clickable screens. The second use is tools of your own. If you do not like the default calculator, you can build one with exactly the features you want as an HTML page. Because you get a “working first draft” so fast, personal tools that solve small problems are easier to try than you might think.

Beyond the chat window: handing the workspace to an agent#

All of the uses above follow the pattern of “ask in the chat window, get an answer, apply it somewhere myself.” You upload a file, copy the response and paste it in, and if it is not right, upload again and ask again. There is a point where this round trip disappears: giving the agent the whole working folder.

Put simply, a developer’s process looks like this: create a folder for the source code, open that folder in a development tool, create and edit files inside it, and check the result in a browser. When a coding agent enters the picture, the agent writes the code, launches the app in a browser to look at the screen, analyzes the logs, and keeps at it until the goal is met. This structure can be applied as-is to non-development work. Set up a working folder where the agent can roam freely, ask it to create the files you need, ask it to edit them the way you want, tell it to learn whatever it does not know how to do (skills), and have it do this and that in the browser.

I first tried this idea with Cursor. Create a folder, put text files in it (notes, requirements, meeting minutes, checklists), open that folder in Cursor, and the agent works while referring to the files in the folder on its own. But Cursor is a development tool by birth, and it had far too many features a non-developer would not understand. Since then, products that offer the same structure for non-development work, such as Claude’s Cowork, have appeared. Point it at a working folder, connect a browser, attach the MCP servers and skills you need, and you can work the same way without being a developer.

Use case: pulling data from an ad console and tidying it up#

Here is one flow I have actually run. The task is to download a targeting list for a particular campaign from an ad console, edit it, and share it.

1. “Open a browser and go to the ad console page.” I log in myself. You could give the agent the username and password and have it type them in, but two-factor authentication and the like are still tricky, so doing it myself is easier.
2. “In the ad console open in the browser, find the targeting list for campaign XXX and export it.” The agent reads the screen, finds the button, clicks it, and runs the download. Register this procedure as a skill and it runs faster and more accurately next time.
3. “Sort the original file by click count in descending order, shade the rows with 100 or more clicks yellow, and save it.” If the original is a CSV, formatting cannot be applied, so the agent converts it to XLSX and saves that.
4. “Upload the Excel file to Google Drive and give me the link.” If a Drive connector (MCP) is attached, it carries on through the upload and the link sharing.

Done chat-style, this would mean four rounds of download, upload, copy, and upload again; inside a “workspace” it becomes a single continuous flow. It is still early days, there are bugs, and sometimes it runs differently from what I intended, but I believe this kind of work process will become commonplace before long.

Choosing by task type#

Summary#

• Chat-style services alone greatly widen what you can handle by yourself: search, summarization, translation, writing, code, data analysis, visual content, and prototyping.
• The next step is to remove the round trips of the chat window. Give the agent a working folder, connect a browser and external services, and pin down repeated procedures as skills.
• Choose between chat-style, your own chatbot, workspace-style, and resident agents depending on the type of task.
• If this is how working alone changes, the next part moves on to building and running a product alone.