Why I, a long-time AI user, have started sharing my insights now
Why I, a long-time AI user, have started sharing my insights now
I have been using AI since its very early days.
I have integrated it into my work, hacked it, and gone through trial and error. By the time those around me were amazed by ChatGPT, I had already entrusted it with part of my daily tasks.
However, I haven't really shared that with the outside world.
This is because I was more focused on "mastering it" than on "sharing know-how." But recently, I started to feel that it's a waste to keep it to myself. I want to properly articulate and share what I've learned through my trial and error. That is the reason I started this note.
For my first article, I will write about "how I have integrated AI agents into my work."
[I knew from the start that AI was "useless"]
When I first started using AI, I noticed something immediately.
It is that "AI is useless on its own."
This is not a criticism. It is just a matter of fact. Even an excellent new hire cannot function if they are not taught anything. The same goes for AI. Unless you tell it what to do, who it is for, and what the quality should be, it will return irrelevant output.
That is why I focused on thinking about "how to master it" from the very beginning.
What I learned through testing was that the output quality of AI is almost proportional to "your own design capability."
[Work design capability was the key to leveraging AI]
If I had to sum up the most important thing for mastering AI in one word, I would say it is "work design capability."
Specifically, it involves these four steps.
1. Set a goal
Clarify what you want to achieve. Be as specific as "deliver three posts a week providing AI agent knowledge for startup executives" rather than just "I want to increase X posts."
2. Break it down into tasks
Break the goal down into units that AI can execute. Think of it as dividing "write an article" into "theme selection -> outline creation -> body generation -> review request."
3. Translate into actionable steps
Define how you want it to work in a way that the AI won't get confused. Who is the reader? What is the tone? What are the forbidden words? What is the output format? The more carefully you do this, the more stable the output will be.
4. Make decisions yourself
Looking at the results executed by the AI and deciding whether to go or not is the human's job. If you try to hand over decision-making to AI, it will inevitably fall apart somewhere. Delegate the execution, but keep the decision-making. This boundary is the most important thing.
As I repeated these four steps, the quality of the output improved.
[What I actually did through trial and error]
I tried various things in parallel while thinking about the design.
First, I created a "specification document" to pass to the AI. I increased the resolution of the target audience and documented it. I clearly specified words to use and words not to use. I accumulated examples of "I like this kind of output, this is not what I want." As I repeated this, the output gradually became closer to "my own words."
Next, I divided the "agents" by task.
I defined the roles, memory, and operating methods for marketing, PM, and CS roles respectively. Instead of leaving everything to one AI, I assigned responsibilities. It is the same way of thinking as a human team. This significantly improved the quality and stability of the tasks.
Also, I stopped seeking perfection.
I don't expect 100-point output from the start. I add 30 points to the 70-point output to complete it. Once you get used to this cycle, the speed of working with AI increases dramatically.
[Tasks currently handled by AI agents at deflag]
Now that I have gone through trial and error, my team has AI agents handling the following tasks.
Every morning at 9:00, a "draft X post for today" arrives on Slack. I just check it and reply "post it." 3 to 5 posts per week are running automatically.
The same goes for drafts of note articles. When I pass a theme, a draft of 3,000 to 5,000 characters comes up. This article is one of them.
When an inquiry arrives, a reply draft is automatically generated. For competitor research, a report comes out just by instructing it to "analyze." Weekly reports are automatically summarized and arrive on Monday mornings.
I believe these work because I did the "design" carefully at the beginning.
In terms of time, about 25 hours a week were saved. 100 hours a month. I am able to use this time for work that only I can do.
[Reason for starting to share]
The trial and error up to this point remained almost only within myself.
However, every time I talked to managers and PMs who also wanted to "master AI" or "wanted to incorporate it into their work but didn't know where to start," I noticed that I was asked the same questions.
In that case, it is better to properly verbalize and put it out there.
In this note, I will write about the trial and error I have gone through and the way of thinking that has emerged from it. I plan to continue talking about concrete topics with the theme of "how to design and use AI."
In the next article, I will write about the basic way of thinking about AI agents and what should be designed first.
Follow me to receive notifications when I post updates. If there is anything you would like to ask or if you want more details on a specific topic, please let me know in the comments.
#AIAgent #Startup #SmallTeamManagement #WorkDesign #OpenClaw