An AI's innate intelligence doesn't change. But here is how changing how I reminded it made it bloom.
That AI might actually be smart, but you might just be "resetting" it every time.
Have you ever thought, "This AI is surprisingly useless"?
About a year ago, I felt exactly that, while actually continuing to reset the AI myself.
In my note from two posts ago (the article below), I wrote a series interpreting articles by HubSpot co-founder Dharmesh Shah, but this time it's a story about a realization that came just before that—one that was subtle but effective.
This AI isn't smart
When I first started using a tool called Claude Code, I often felt, "This AI isn't actually that smart."
It answered so accurately yesterday, but today it's like we're not even on the same page. Even after explaining the project's premises, the moment I opened my computer again, it was back to square one.
At the time, I thought it was a lack of ability on the AI's part. I thought, "So this is all this model can do." But looking back now, the cause wasn't the model's intelligence, but my own way of using it.
I would rewrite the same explanation from scratch every time. If you add that up over a week, I think I spent about 7 to 8 hours just on this "explanation of premises."
That's a full day. I was supposed to have started using AI to make my work more efficient, but I ended up losing a day just explaining things to the AI.
The story goes back to around February 2025, when Claude Code had just been released.
At the time, I was using Claude Code in the CLI (that environment with just text on a black screen where you operate software via command line).
Moreover, the settings were placed in the largest possible scope: the entire computer. This meant it wasn't segmented by project.
The premise that AI is stateless (doesn't remember context)
I understand it now, but AI basically has a property called "stateless". It's a difficult term, but you don't need to memorize it.
In a nutshell, it's the property of "forgetting the context up to that point completely the moment the conversation ends".
If it were a human, saying "About that thing we talked about yesterday" would be enough. But AI doesn't have that sense.
I didn't know about this stateless property back then. So, every time, I would input the project background and my requests from scratch.
It felt like introducing myself to an elderly person with memory loss every time, saying, "Nice to meet you, this project is about this, and I'm receiving this kind of request," before finally getting to the actual task.
Explaining the content and the actual task—it was always double the work. And since the elderly person wouldn't remember the previous conversation at all, if I cut corners, I'd get an off-base answer. So I couldn't cut corners. It was this cycle over and over.
A turning point arrived for me, someone who was spending 7 to 8 hours a week—a full day's worth—just on these 'overview explanations'.
'Why does the premise disappear every single time?' It was a simple question.
How can I make it remember the context?
As I researched, I arrived at a mechanism called CLAUDE.md. This is a feature officially provided by Anthropic, where if you place a file named CLAUDE.md in your project folder, Claude Code will automatically read the contents of that file every time it starts up.
In other words, it means you can write the project introduction part, such as 'here is the background,' into a file in advance and provide it.
Since learning about this mechanism, I have made it a point to create a CLAUDE.md for every project.
By the way, if you type the slash command /init, it will create it for you automatically.
As a result, from the moment I wrote that file, Claude Code began to converse with me as if it had been involved in this project all along.
The effort of explaining the premise vanished, and I could talk about the main topic right away. I was able to start the conversation from where we left off, rather than from zero.
At this moment, I remembered a sentence written in a newsletter by Dharmesh Shah, co-founder of HubSpot (a person who continues to actively share information about AI utilization today).
The model's IQ doesn't go up, but
'The model's IQ doesn't go up. But its EQ does. That's because it becomes able to understand the user more deeply.' This is the gist of his point. *Quoted from the article below
In this article, Dharmesh explains how to engage with AI as a 6-level 'ladder'.
6 levels for using AI smartly
Prompts: What to ask each time
Custom Instructions: Reusable self-introductions and premise information
Skills: 'Patterns' for specific tasks (sets of instructions, templates, and examples)
Plugins: Expansion of functionality itself
Tools: The power to actually execute something
MCP, API, CLI: Mechanisms for connecting to external systems
This is exactly what I experienced. Claude Code itself (the model's raw intelligence) might not actually have become that different from a year ago to now.
What changed is that I became able to properly pass the context of 'premises' to the AI. By simply adding another layer of 'custom instructions' or 'project memory' on top of the prompt (the instructions for the AI), the same model started to look as smart as a different person.
If you currently feel like 'the AI just doesn't get it' or 'I feel like I'm explaining the same thing every time,' you don't need to build a large-scale system first.
The first step is to write just three lines about your project, even in a notepad app. For example, it looks like this:
1. Customer acquisition support project for the XX industry
2. For sole proprietors in their 30s
3. Please explain without using technical jargon
Just by passing these three lines along the next time you talk to the AI, the experience should change significantly. Once you get used to it, if you are using Claude Code, you can save it directly under the project root with the filename CLAUDE.md, and you won't even have the trouble of passing it from next time.
Have you ever had the experience where you thought, 'This AI doesn't understand me,' but in reality, you were just making it forget the explanation every time?
Please let me know in the comments if you'd like.
At another time, I plan to write about how I climbed from the third rung of this six-rung ladder, 'Skills,' and above (and where I stumbled).