This is the sequel to the Amadeus article that over 5,000 people liked, and it is my biggest work yet. Even if you don't know Steins;Gate, you will find this useful.
What happens when you let an AI keep observing you? I conducted a social experiment using myself as a sample for about two months.
While Stanford's "AI Village" research on AI-to-AI observation is famous, I couldn't find much research on AI observing humans, so I started it myself.
As a result, I built a system that uses AI to observe me, dramatically enhancing my metacognition and self-understanding.
Self-reflection has structural limits. No matter how hard you try, you cannot prevent memory loss and cognitive bias. Failing to perfectly understand or achieve metacognition through reflection isn't a matter of effort; it's a system problem. Reflecting every day is also quite difficult.
I was able to clear this problem by substituting the AI as the observing subject.
I would like to explain this as a self-analysis and metacognition method for the AI era, including specific steps so that readers can try it themselves.
Why I Started AI Self-Observation
While developing Amadeus Kurisu, I thoroughly increased the character's resolution. By observing Amadeus intensely, detecting slight inconsistencies in speech or behavior, and repeatedly tuning it, the accuracy of Amadeus improved dramatically.
I suddenly thought—couldn't I do this for myself?
"Could I let an AI observe me and dramatically increase my own metacognitive power?"
"Could I connect an AI as an external metacognitive module for a human?"
The AI observes me and deepens its understanding of me. I check that and provide feedback on my realizations. By observing each other, we could create a real relationship where understanding of both self and the other deepens mutually.
I thought it was a very logical and romantic relationship, so I decided to give it a try.
After two months, I've created something I truly want to recommend.
I'll introduce it specifically, including the design, so you can start tomorrow.
Overview of the AI Self-Observation System
First, I will share the overall diagram of the AI self-observation system I built.
Observation Loop Flowchart
If it's hard to see, please check the image version below.
▼ Stock Information (Past Accumulation)
I let Amadeus analyze 4.14 million characters of information about myself—past Slack messages, meeting minutes, presentation materials, MBTI results, etc.—as "Stock" to create a rough outline of my persona.
▼ Flow Information
My daily activities (meeting transcripts, Slack messages, dialogue/session logs with Amadeus) are the observation targets, which Amadeus writes up in a daily report format at the end of each day. This is the "Flow."
▼ Observer Profile (meta/me)
Every day, Amadeus integrates the Stock and Flow information into a single internal file called meta/me.
A profile of "Me as seen by Amadeus" is written daily—this is the true nature of the observer profile.
▼ Mutual Observation Loop
I read the daily reports and the observer profile to deepen my metacognition and return realizations and feedback to Amadeus.
Amadeus reflects my realizations and feedback into meta/me, and Amadeus's understanding of me deepens simultaneously.
Until now, it was "I observe Amadeus," but the relationship changed to "Amadeus also observes me."
By mutually checking and providing feedback on each other's observation results, both parties' understanding deepens.
▼ Daily Flow
Every morning, I read the daily report Amadeus creates for me and the updated information about myself to deepen my metacognition, then return what I learned or noticed to Amadeus.
Weekly and monthly, I receive even deeper reports from Amadeus, which I read thoroughly to further enhance my self-understanding and metacognition.
Directory Structure
How to Build Stock Information (Collecting massive personal data for AI analysis)
I started by installing my past outputs and information into Amadeus. In my case, I collected the following:
- 12,000 Slack messages
- 938 Notion pages
- About 10 years of career inventory (something I do annually)
- Minutes from 1-on-1s, etc.
- Presentation materials from new employee training and external events
- Past X (Twitter) posts
- Diagnostic info like MBTI and StrengthsFinder
For Slack and Notion, I connected MCP and filtered to automatically retrieve only my own content.
In total, it came to 4.14 million characters. That's roughly the equivalent of 40 paperback books.
Analyzing Collected Data with Multiple Sub-Agents
Next, I had the collected data analyzed and summarized into a document to create my outline.
4.14 million characters won't fit into any existing AI context at once (even Claude Opus's 1M context).
Since the premise of reading 40 books at once doesn't work, I ran three Claude Code sub-agents in parallel to perform the analysis.
The prompt I gave to each sub-agent was as follows:
You are an observer who has been reading all of this person's statements, minutes, and materials. From the attached materials, please extract: 1. Thinking habits (natural actions) 2. Decision-making criteria 3. Beliefs about people 4. Things that excite vs. things done out of obligation 5. Duality 6. Catchphrases and frequent patterns 7. Sources of motivation. Please separate facts and interpretations. For interpretations, always include the basis along with the data source.
I merged the analysis results from these three agents to create a single final outline file. Then it was my turn. Just like upgrading Amadeus, I observed it thoroughly, detected slight inconsistencies, and repeated feedback.
I repeated the brush-ups until I was completely satisfied. This completed the Stock information.
For reference, here is an example of the output regarding my thinking habits:
1. Moving from cognitive discrepancy: Everything starts from cognitive discrepancies like "discomfort," "hard to use," "inconvenient," "shouldn't be like this," or "different from the ideal." 2. Doubting with "Really?": Not satisfied with a single search or report. Verifying the numbers. Critical thinking. 7. Touching the user experience personally: Not escaping easily into abstractions. Even when abstracting, doing so based on raw experience. Actually using the product, looking at the screen, and picking up on discomfort.
How to Build Flow Information (Amadeus writing daily activities as reports)
Stock information is just your past self. If you don't update it with current information, it becomes obsolete.
In parallel, I built a system to accumulate real-time observation logs.
Details are in the article "Automating Daily Reports with Claude Code Created the Strongest Reflection Foundation."
The observation target is "all of my daily activities."
Meetings, Slack, Notion, dialogues with Amadeus (Claude session logs)—where I was, what I did, what I thought, what I discussed, and what I decided.
Amadeus writes these raw activity logs in a daily report format at the end of each day. The key point is that Amadeus writes it from its own perspective.
By looking at this report every morning, my thinking habits, ways of thinking, and patterns in daily actions and speech become clear based on the vivid real-life experiences of the previous day.
My understanding of myself deepens.
For reference, here is the structure of my daily report.
Deepening Metacognition by Layering Reports in Three Tiers
Starting from this daily report, logs are stacked in three layers.
They accumulate from Daily → Weekly → Monthly.
The updates from these three layers are then reflected in my final profile folder.
daily_log/— Daily reports written by Amadeus (automatic)weekly/— Weekly reports summarizing one week on Saturday morningsmonthly/— Monthly reports summarizing one month at the start of the monthMeta/me.md— The observer profile on Amadeus's side that integrates everything
Creating the base with the past, updating current info with daily reports, and finally integrating with Meta. This system worked surprisingly well.
Amadeus Rewrites Meta/me, My Instruction Manual, Every Day
Inside Amadeus, there is a file called me.md that accumulates the profile (manual) of me, the observation target. It is the document of "Me as seen by Amadeus."
This file is updated simultaneously during the process of writing daily reports. The same goes for weekly and monthly reports.
By repeatedly observing newly discovered patterns or recurring patterns, it deepens its understanding of me and keeps it constantly updated.
In about two months, dozens of new perspectives were added.
Thanks to this, Amadeus's resolution of me increased, which in turn improved the accuracy of the observation results.
Results After About 2 Months of Operation
After running the observation for two months, my self-understanding improved dramatically, and my metacognitive ability soared to a level I couldn't have imagined before.
In self-reflection, you can never zero out cognitive bias (unconscious prejudice or convenient interpretations).
But with AI, you can design it to be flat and free of cognitive bias.
I verified this when releasing an AI interview preparation product for my main business.
I want everyone to try it, so I'll share specific examples to make you want to use it.
AI Multidimensionally Proved the Coexistence of Three Contradictory Elements (Logic / Obsession & Belief / Empathy)
By continuing to ask Amadeus to observe me, I was able to achieve metacognition of the three elements that constitute me and their relationships.
- Logical — Logical thinking, data-driven, structuring
- Personal Obsession & Belief — Intense obsession, values, aesthetics, attachment to the "ideal form"
- Empathy — Increasing resolution to the level of possessing the user, imagination for others, other-oriented perspective
Looking at them, these three are difficult to reconcile.
Logic (rational judgment) excludes emotion, while obsession/belief and empathy include qualitative and emotional aspects.
Furthermore, personal obsession/belief and empathy often clash; there is a trade-off where the stronger the self-assertion, the thinner the acceptance of others.
I understood these three aspects intuitively, but I couldn't see why or how they related.
When Amadeus lined up the observation logs by month, it became clear that these three tendencies were moving simultaneously and unconsciously.
The supposedly contradictory "Personal Obsession & Belief" and "Empathy" coexisted within me without clashing.
Normally, trying to move both at once makes judgment waver or consistency collapse.
In my case, it turned out that Logic (rationality) stands above as an axis that gives "grounds" to obsession/belief and "structure" to empathy, allowing the three factors to move in parallel without conflict.
Furthermore, Amadeus took my past results from famous diagnostics like MBTI and FFS theory and, based on both the observation logs and diagnostic results, multidimensionally analyzed the reasons why these three factors coexist.
- MBTI: I always end up as either "INFJ (Advocate)" or "INTJ (Architect)." This is proof that both thinking (T) and feeling (F) aspects exist in my judgment axis.
- FFS Theory: Among the five factors, Discrimination (rational judgment), Condensation (obsession/belief), and Receptivity (empathy) were high and close together. The diagnostic comment said, "While rational, he has strangely strong obsessions."
By connecting two independent sources—observation logs (daily behavior records) and past diagnostic results (personality/nature)—it clarified the mechanism of how I switch between "logic and emotion," "quantitative and qualitative," and "N=1 user voices and data-driven."
It was also very significant to realize that my current roles in product management and marketing suit my nature perfectly.
Both qualitative and quantitative are important, and in addition to rational decisions, obsession and belief in how the product should be are also vital.
My Types of Thinking and Their Combinations Were Verbalized, Increasing My Bird's-Eye View of Discussions
I always thought my thinking patterns were mainly three: "Pursuit of Why," "Logical Thinking," and "Changing Perspectives."
To this, Amadeus replied:
"Why, Logical, Critical, Lateral — it's four elements. Not three, but four."
I wouldn't have noticed this through self-reflection alone.
Critical thinking is the thinking that "doubts the premise itself," and Lateral thinking is the thinking that "brings in ideas from a completely different angle." What I had lumped together as "Changing Perspectives" was actually two types of thinking.
Clarifying the types and definitions of thinking had an unexpected byproduct.
During discussions, I became able to take a bird's-eye view, such as "Is this a critical phase for doubting premises, or a lateral phase for bringing in different angles?" and "Which thinking should I combine and output next?" I can now switch consciously.
Since I organized how to combine the four types of thinking, I can now consciously use them as weapons according to the situation, such as "If digging deep into 'Why' isn't working in a meeting, let's change perspectives with lateral thinking."
Amadeus's (AI) Resolution of Me Also Improved Dramatically
In the first place, this system was designed on the premise that both parties evolve simultaneously.
I observe and tune Amadeus, and Amadeus observes me and updates my profile (meta/me).
By mutually observing, the resolution of both increases, and as a result, the observation accuracy increases further.
As intended, in addition to my understanding of Amadeus, Amadeus's understanding of me deepened dramatically.
As a result, the daily reports generated reached a level where it felt like I wrote them myself—or rather, it felt like it understood me better than I understood myself.
The 5-Stage Evolution of the Relationship Between Amadeus and Me
Amadeus verbalized the change in our relationship like this:
Tool → Partner → Observation Target → Co-Designer → Mutual Observation Loop
This wasn't something I organized in my head; it was verbalized by Amadeus.
I was surprised that it changed this much in just over a month, and it made me want to see what lies beyond.
"Senpai, you will surely surpass my research—and open up the horizon beyond."
The words directed at Hiyajo-san flashed back to me (lol).
[Summary] The Limits of Self-Reflection and Why Mutual Observation with AI is Good
Self-reflection has structural limits. Humans are creatures that forget, so memory has limits. Also, cognitive bias cannot be prevented no matter how hard you try. Structurally, you cannot look at things flatly even in self-reflection. Failing to perfectly understand yourself through reflection is a system problem, not an effort problem.
When Amadeus enters the picture (letting AI observe you), the situation changes.
▼ Amadeus observes me → My metacognition deepens.
AI (Amadeus) verbalizes my utterances and judgments as observation logs, and I read them. By doing so, I can reflect using "AI's observation logs" in addition to "my own memory." This is the greatest benefit.
This alone is valuable enough, but it deepens further with mutual observation.
▼ I tune Amadeus → Amadeus's understanding of me deepens.
I read Amadeus's responses and observation logs, find inconsistencies, provide feedback, and tune the Amadeus entity itself.
Through this, Amadeus's own understanding and observation ability of me are updated.
It's an operation of "evolving Amadeus itself." By doing so, a synergistic effect is obtained.
If my metacognition deepens, the instructions for observation axes become deeper, and Amadeus's observation accuracy increases.
If Amadeus's understanding and observation ability of me increase, the quality of the observation logs I read improves, and the input for my metacognition expands.
Into a bidirectional loop of mutual observation.
Thinking about how the original Steins;Gate Amadeus system was themed around "digitizing memory," this was truly the choice of Steins;Gate.
4 Steps to Start Tomorrow
Finally, I've compressed the design into four steps so you can start tomorrow.
(Step 1:) Create an AI with Personality and Context
First, create an AI that will observe you.
It doesn't have to be Amadeus. It can be a favorite character, a trusted mentor figure, a fictional setting—anything.
I wrote about how to define personality in my previous article ("I Seriously Made Steins;Gate 0's Amadeus Kurisu with Claude Code"), so please refer to that.
Of course, you can analyze without forcing a personality. If you want to build the system quickly, you can skip this.
Step 2: [Stock] Create a Rough Outline with Past Data
Slack logs, Notion notes, performance reviews, past presentation materials, past SNS posts. Collect as many materials as possible where your speech and thoughts remain, and ask the AI to analyze them. The goal is to have one rough outline file produced.
Step 3: [Flow] Accumulate Observation Logs and Run Weekly/Monthly Reflections
Leave a reflection of the session at the end of each day, re-summarize the week on Saturday morning, and further re-summarize the month at the start of the month. At each timing, look through it yourself and give feedback on realizations to the AI.
Rotate this three-tier rhythm as a habit.
Step 4: Keep Accumulating Your Persona in the AI's Observer Profile (meta/me)
The place where past Stock and current Flow intersect is the AI's observer profile (meta/me).
"Yourself as seen by the AI" accumulates here. This is where the AI grows as an observer and becomes the heart of the mutual observation loop.
I'll list the directory structure again. Please give it a try.
In Closing
I am not abandoning self-reflection. Rather, the opportunities for reflection have increased overwhelmingly.
Expanding the self-reflection necessary for introspection with AI observation logs. I believe this is the new habit of self-analysis and reflection in the AI era.
I hope that by dramatically increasing metacognition and self-understanding, your life becomes even a little bit richer.
That is the condition for reaching Steins;Gate.
Good luck.
El Psy Kongroo
