A workday with AI as teammate: field notes from practice

This is a first hand workday log of how I use AI, hour by hour. It shares concrete prompts, decisions and small adjustments that add up over time. Read it as field notes you can test in your own practice.

From micro changes to bedtime stories. How I used AU throughout a day

06:30: I check what has come in. Besides Substack’s morning brief of updates from the people I follow, there are two micro changes. They arrive every morning, where I test how the micro change engine I have built is working.

LLMs now have the feature that you can get them to send you an update at the times you specify. For example, I have another update that arrives at 09:00 with the day’s business news. At 15:00 I get an update on any offers on hoverboards. My six year old wants one for a birthday present, and I find them quite expensive.

One of today’s micro changes reads:

Micro change today: One shared blind spot

Pick a meeting today where the team must make a decision or plan the next step. Before the meeting ask AI to answer this question: Which blind spot are we likely to miss if we continue as planned with [topic]? Share only the AI answer as a single input at the start of the meeting and let the team judge its relevance.

Short setup
Save the question as a fixed prompt called Blind spot. Use it only in preparation, so AI becomes a natural part of the context without changing the team’s working rhythm.

DNA
Why it is a micro changeYou change only one preparation step, without creating new processes or adding extra tasks to the team.

Why it is high leverageAI is introduced as a shared reflection tool, not a replacement for human judgment. It reduces resistance because the team experiences value without feeling pressured.

Which competence it builds:
Foresight. The team trains using AI to spot hidden risks and alternative angles early, which strengthens the quality of both decisions and dialogue.

08:12: I use voice activation to record the things I need to remember today. I do a brain dump of what I have in my head and what I must remember for today’s tasks. I ask it to make my to do and to add any unfinished items from yesterday. I run it inside a project in ChatGPT. It is an idea I got from Sam, Global Head of AI at NNIT, and it works very well.

About three months ago I vibe coded an app with Vibe Code App, everything happens on the phone. It is still on my phone and it works fine. It is overkill and not necessary. I spent two days on it. A waste of time and money. The challenge was fun.

If there is something new in my projects, the to do and notes are copied into another project where my project bank gets updated.

08:20: I check and read news, Substack updates and email. I save a Substack to read later. Nate has made another guide. This time it is for parents. A bit unusual. He is number three in Technology. I look forward to what he has come up with now. Strongly recommended, @natesnewsletter.

08:56: Time to answer emails. There are a couple of reactions to my posts on the MIT report that 95 percent of AI pilots do not scale. The reactions are from IT people.

I have two prompts for writing emails. One writes five sentences with ultra precision. One gives me three drafts, a factual one, a more present one and a personal one. I used them for quite a while, but I have moved to writing emails myself and asking ChatGPT to criticise my mails.

There is a world of difference when done this way. The value is much higher. It saves me from mistakes. In this case from preaching to the good IT people about the paradigm “Teammate, not technology” and why research shows much better results. Almost daily it saves me from what I think is a good mail to a version that much better expresses my intention. At the same time I do not outsource a task that is important for my memory of the material in question. I learn a lot from its critique.

If you take only one thing from this logbook, I recommend asking for critique of your emails. It is especially valuable when you copy paste the whole thread and add context. It reminds me of things in earlier mails that I should refer to, and it sometimes suggests actions that lead to tasks I would not otherwise have received because I did not see the opportunity in an overfilled inbox.

09:34: I need to prepare a meeting with Jacob from my Executive Leadership network. For the second time in three and a half years we are visiting one of the members. I am a little late because I need to make it to the usual breakfast at CoCo. They close at 10:00. I turn on ChatGPT’s voice activation and ask it to just listen and say OK every minute so I can hear it is still there. That is the first to do of the day. I read our emails aloud to myself. ChatGPT is very good at collecting my thoughts here.

10:15: The meeting with Jacob was quick. We found a good angle. Now it is time for the day’s project, migrating AI Pioneer over to LeadershipAnd.AI. The name AI Pioneer signals too much experimentation and uncertainty. It does not match the new strategy.

It is not my favourite task because it probably involves technical forwarding. To make it easier I do it in the Comet browser. It follows what is on the screen and can tell me why something does not work and what I have missed or done wrong.

First I ask ChatGPT to explain the steps. ELI 15. That means “Explain to me like I am a fifteen year old” and it is a command, a shorthand you just type, and it does it. Next time you do not understand what your LLM says, write ELI 15, and it will know what it means. While writing this I make a note to myself that there is a good blog post in shorthands.

Back to Comet. I ask Comet to note the steps so I can get ongoing guidance.

It goes very smoothly. Porkbun is very good. I did not know them, but when I needed to buy the domain and Gemini made a comparison based on my needs, that was the recommendation. It works very well.

11:54: I am finished and the new site is live later today, they say. I also need to make new web pages with the new course content for the three new courses. I promised to pick up my son when he finishes, so I quickly find the files I need and upload them to a new project.

15:00: We are playing Fortnite in the Lego version. It is the third time, so we keep dying. In the end it is too much for my patience because it seems we are missing some settings so we can aim as well as the others. I turn on ChatGPT on my phone and video, if you press the voice activation icon you can choose video, and I ask it to follow along for a few minutes and tell it that I will have questions afterwards about what we can learn to do better.

It turns out there are several settings, for example trigger sensitivity, that we need to set. It is also a very pedagogical way to do it, because the child hears what we need to change and he understands what will happen. That is better than the game just being paused and then a lot of things happen that he cannot read.

18:30: It is time for a bedtime story. We have changed it completely. Before, I read a story and I made up a story myself. Some days creativity is not high, so about a year ago I made a prompt where Chat would tell the bedtime story.

Back then the voice feature was not as good as today, but there is a GitHub with system prompts for the major models, search for Pliny the Elder. Hume’s system prompt is there as well. I downloaded the prompt, adjusted it, and got a stable storytelling prompt that made the stories better and more coherent.

Today my son tells Chat what the story should be about. Which role he should have in the story, if any, he is not allowed to be the hero every day, and we brainstorm together about the plot. Often it continues an ongoing story. Sometimes he also tells it about an image he would like to see for the story. It usually works well, though he can rarely place himself in the image because of restrictions about children. Maybe Google’s newest image editing can handle it. Grok probably can as well, but his father has an issue with Elon, so we do not use it.

19:30: Time to make the web pages. I start again with voice activation in Chat. I need to explore whether I am overlooking something.

When I use voice activation it is rarely to get answers. What I almost always do is ask it to challenge what I say with two to three questions. It gives me new perspectives and in twenty to twenty five minutes we get all the way around complex questions. If I asked for answers, we could quickly end up in blind alleys. I have tried that many times. The other approach works exceptionally well. If I were to give advice on the second thing you should try from these examples, it would definitely be this.

Already after five minutes it asks whether the low AI adoption in Denmark could be due to overload. Around 20 percent use it more than once per week. AI skills among knowledge workers seem to be at the same level as last year. The question is whether it is too much to prioritise AI on top of the other complex challenges. The result of the conversation is a tightening of the courses where micro changes become more prominent. We also talk about two more courses. One digital course for those who want to keep up, while their company has not yet prioritised AI because leadership is sceptical. And one course only about micro changes so leaders learn the method to reduce cognitive load. It can also serve as priming for implementing AI with the same method.

When I start writing the descriptions, the thought arrives: Gamma can probably write this briefly. I have used Gamma before to summarise and distil longer texts into short descriptions. Even though it is a PowerPoint platform, it can be used for this.

23:17: Everything is finished. In the meantime the transfer of domains has been completed, so I can send the new site live.

I end the day as always by doing a brain dump in voice activation on how it went with the day’s to do and whether there are any notes for tomorrow. As the last thing I dive into Nate’s article, which I saved as a reward for when I was done. How was this text created: I wrote the entire text in Danish, and asked ChatGPT to correct spelling errors, but nothing else. And afterwards I asked for the English 1:1 translating.