AI Project Management and Owning the Context
Starting with calorie tracking
I want to start with a very simple example. For about a year, I wanted to start tracking my calories. I even installed the FatSecret app because it was the app most people recommended at that time. But I never actually started using it. It simply felt too complicated. Every meal required weighing ingredients, measuring portions, and entering everything manually. It looked like something only people with a lot of free time and endless discipline could do consistently. I tried several times but never managed to build the habit.
Then I realised that I did not actually need another calorie-tracking application. I wanted to see whether I could do the same thing using my everyday LLM. It did not matter whether it was ChatGPT, Claude, or Gemini. So I started with ChatGPT. I simply took a photo of my meal and added a short description. Buckwheat, chicken, salad, pumpkin seeds, olive oil and salt. ChatGPT estimated the portions, calculated the calories and macros, and gradually built my nutrition log. It was simple enough that I actually kept doing it.
One chat for one purpose
Very quickly this became much more than calorie tracking. Every meal, every photo, and every calculation stayed inside the same chat. Because all of that context remained in one place, I could ask for any report I wanted. Calories for today, protein for the last week, average nutrition or trends over time.
The same idea works for many other tasks. I have another chat that is used only for translation. I gave it one instruction at the beginning, and now everything I write is translated in the way I want. Instead of starting from zero every time, I use different chats for different purposes. Every chat builds its own context and becomes better at its specific job.
From chats to projects
The next thing I discovered was Projects in ChatGPT. For a long time, I did not understand why they existed, so I ignored them. Once I started using them, their purpose became obvious.
A project is simply a layer above individual chats. It can be a private project under NDA or just a normal project for everyday work. Inside one project, you can have as many chats as you need. Each chat might cover a task or mini project. For example, one chat within the project for information architecture, another for discussions, meeting transcripts, brainstorming sessions, documentation, implementation questions, or anything else related to the same product.
Instead of having dozens of disconnected conversations, everything stays together inside one project. Over time, the LLM understands more and more about the project and product because it keeps all of the context in one place.
Context is the real value
Software engineers have already gone much further with this idea. Their AI projects can be connected to GitHub, Jira, documentation, and many other tools.
For me, as a product designer, the most important thing is not the infrastructure. The most important thing is understanding and owning the full context.
I want my project to know everything that has happened from the very beginning. Discussions, meeting notes, research, design iterations, problems, mistakes, and decisions. The earlier I start working inside a project in my LLM, the more valuable it becomes because it continuously collects context.
After some time, it starts to feel less like an AI chat and more like a colleague who has been working with me on the same product for a year. I can ask whether we solved a similar problem before or whether we can reuse an old solution. It already knows what I am talking about. For me, this accumulated context is the most valuable thing an LLM can have.
Example one. Discovery
Let me give you a practical example. Imagine a company asks me to improve an existing digital product. At the beginning, I know nothing about the product because it is completely new to me. It could be a website, a portal, or an enterprise application. The goal is simply to make it better.
If I have seven working hours for this task, I will probably spend six of them on discovery. Generating ideas is easier than identifying problems and pain points. Understanding the product is the difficult part. I need to understand the business, the users, the user journeys, the strengths, the weaknesses, and why previous decisions were made.
Today I do not do that alone anymore. I explore the product together with my LLM. As I learn something, I immediately share screenshots, notes, observations, and questions. The AI is not generating solutions yet. It is learning the product together with me.
This is especially important for private enterprise products that are not available on the public internet. The AI cannot simply visit the website and understand everything by itself. I have to explore it.
Traditionally, all of that knowledge stayed only inside my head. Today I build that understanding inside my project with an LLM at the same time.
When discovery is finished, both of us understand the product. I understand it, and my LLM understands it as well. Then I can ask it to identify problems, suggest improvements, challenge my ideas, or generate ten different solutions. The quality of those ideas comes from the shared context that we built together.
AI project manager
This is where I see a similarity with a traditional project manager. In many companies, there is a person who attends meetings, joins workshops, takes part in user interviews, reads documentation, and gradually becomes the person who knows everything about the project. When I say “AI Project Manager”, I simply mean a project folder within my LLM.
My AI project manager plays a very similar role. The difference is that it never misses a meeting, never forgets a decision, never gets sick, and never goes on holiday. Every meeting transcript, workshop, research session, and document becomes part of the project’s knowledge. Over time, it becomes a living internal knowledge base that understands not only what happened but also why decisions were made.
In other words, it becomes the project’s internal wiki.
This becomes even more useful if the project receives every meeting transcript and is connected to tools such as Slack and the rest of the project documentation. Then it can know almost everything that has happened. As a designer, I can ask why a screen should work one way rather than another. The LLM can tell me that a client or manager explained the reason in an earlier meeting and can show the logic behind the decision. In the past, I would ask a project manager whether option one or option two was better. Now I can also ask my AI project manager. It is always available and answers with the full project context.
Example two. Designing with context
Here is another example. I am already working inside an LLM project that understands the product and all of its context. A new requirement arrives. We need to design another dashboard screen that displays reports across different regions, dates, and other filters.
The requirements document is long and detailed. Instead of analysing everything alone, I upload it into my project and ask one simple question. Do you understand what needs to be built? Usually the answer is yes because the project already understands the product.
Then I ask another question.
Based on everything you know about this product, suggest the simplest and most usable solution, then generate prompts for Claude Design and Google Stitch to design it.
The important point is not prompt writing. The important point is that those prompts are created by an LLM that already understands the product, the business, and the users.
Before using those prompts, I also provide around ten existing screens from the product to Claude Design or Google Stitch. Sometimes there is no proper design system yet, but there are already wireframes or UI screens. I ask Claude Design and Google Stitch to continue the existing style instead of inventing a completely new one.
Instead of generating random UI, Claude Design and Google Stitch produce design mockups that are relevant to the existing product.
Then I simply run both tools. Within a short time, I receive two design options. They are not final designs. They are the first proposals that start the discussion.
I show both versions to the stakeholder and ask a simple question. Is this the direction you wanted? The answer is yes. Then I ask which version they prefer. In this case, they selected the concept generated by Claude Design.
The biggest surprise is not which tool produced the better design. The biggest surprise is the speed. In less than an hour, I can go from a written requirements document to two realistic design mockups for a new feature that already fit the existing product.
For me, this is the real game changer. AI is not replacing design. AI is helping me reach the first meaningful solution much faster because it already understands the project. And please, always remember: we use AI for speed, not perfection.
Build context continuously
Today I try to work inside LLM Projects instead of isolated chats. Every new design, meeting outcome, requirement, mistake, and lesson learned becomes part of the project knowledge. I continuously update the project so the LLM grows together with me.
The more context it collects, the more useful it becomes. Not because the model changes, but because its understanding of that specific product becomes deeper.
Managing context has become just as important as managing the project itself.
The current limitation
There is one important limitation today. Projects in ChatGPT are still personal. The accumulated knowledge belongs only to my account. I am the only person who benefits from this growing project knowledge.
Ideally, this context should be shared across the whole team. Product managers, designers, developers, researchers, and other stakeholders should all work with the same AI assistant that understands the same project.
Today this is not possible. Projects are still individual instead of collaborative. I believe this is the next big step. When AI projects become truly shared across teams, AI project management will move to a completely different level.