SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

[Part 2] The Story of How My Claude Code 'AI Secretary' Grew into 5 Departments in One Month‼️

異世界の音楽家｜ai motion musics

Hello! I'm instkoni, an AI creator.

I usually run a YouTube channel called "Musician from Another World / AI Motion Musics".




[You can also listen to the article summary on our Podcast‼️]


This article is the second part.



In the first part, I wrote about "the limitations of a single chat AI", "setting up a secretary in 15 minutes", and "the experience of assets accumulating like an RPG after one month of operation".



The " alter ego " I raised in the
first part grew to a scale that couldn't be handled by one person in a month is the story of this second part.



In RPG terms, this is the "phase where I transitioned from solo play to party formation." This second part focuses entirely on organizational theory🎯



Specifically, it covers the scene where it naturally expanded from just a secretary's office to 5 departments.
Also, the daily report flow, and the script x AI agent feedback loop.




This is the reality of operations that even a side-hustler can manage. If you haven't read the first part, please check it out here👇



The reference source is the same as in the first part: the video by Shin@Programming Tutorial titled "[Solo CEO] Let's build a company with Claude Code and delegate work to AI employees!"




This article is not the full-set version by Shin, but rather a log of my one-month journey following the route recommended by Shin himself: "
Start with a secretary and expand departments"📘



I'll show you the departments that ended up being created first.




"Spontaneous emergence of departments," "Information coordination rules and daily reporting flow," and "PDCA loops for management scripts" will be discussed in this article. Let's get started.







Chapter 4: The Spontaneous Emergence of Operational Departments — From a Secretary's Office to 5 Departments and Over 40 Scripts🌐

One month ago, the only department in my Claude Code environment was the Secretary's Office.




Yes, this is my pseudo-company as of April 28, 2026, and this is how it looks now.




The departments are located under .company.





📗Secretary's Office: Point of contact, TODO management, and brainstorming
📗Editorial Department: Note article planning, production, and quality control
📗SNS Operations Department: SNS rewriting and distribution after Note article publication
📗Automation/QA: Script maintenance, incident management, and dashboard management
📗Shared Assets: Shared prompts and asset ledger for all departments

Basically, I check the ToDos with the secretary, and depending on the task to be performed, I have them connect me to the agents in each department, which allows me to cycle through 'planning -> implementation -> testing -> feedback'.




Is the text a bit too small?



🟢 The 'Same Task Twice' rule worked


The reason the number of departments increased is clear.



Secretary Office CLAUDE.md contains the rule: "If a task in the same domain is repeated two or more times, propose creating a department," and it worked exactly as intended.



As I introduced in the first part, this rule is based on a comment Shin (@Shin_Engineer) made on X:



"Personally, I think it's more educational and efficient to start by creating a Claude Code secretary, and create and expand departments only when they become necessary."

— Shin@Programming Tutorial




"Create departments when they become necessary." Implementing this single phrase into CLAUDE.md resulted in a 5-department structure after one month✅



Here is how it actually progressed.

📘 Timeline until the birth of 5 departments👇
Late March: Requests to 'write Note articles' started trickling into the secretary's office
Early April: Consultations on article planning and pipeline operations occurred 2-3 times a week
　　→ The secretary proposed, 'Why don't we create an editorial department?' → Immediately adopted
Mid-April: Consultations on distribution and rewriting after Note article publication increased
　　　　　+ Requests for script debugging increased
　　→ SNS Operations Department & Automation/QA were launched one after another
Late April: Shared management of prompts and assets became necessary.
　　→Shared Assets department was separated

The key is that the secretary 'proactively proposes the creation of departments'.




It's something like this. SNS operations are running at a support level, but
everything else is fully operational.



You don't need to think about 'creating a department' yourself.



'(Understands the situation and makes proposals)'



This is actually a big deal for side-hustlers.
Because the secretary removes the bottlenecks in your thinking👀



※ Of course, you can also use them as a sounding board to discuss creating a department yourself.



The mindset of turning repetitive tasks into systems is truly important.



🟣 The Reality of the Automation/QA Department — Managing Over 40 Scripts

The Automation/QA department is particularly interesting, so I'll dive a bit deeper.



This department currently manages over 40 Python scripts.

I've blurred the details a bit since some of them can't be shown (lol).




There are about 30 scripts under `scripts/` in the main repository, and if you include the separate repository `sns-content-automation/note-content-automation/`, it totals over 40.

📗 Here is the breakdown👇
①Pipeline-related (draft_writer / proofreader / meta_generator / html_converter / note_publisher / podcast_writer, etc.)
②Editing support-related (infographic_selector / news_collector_jp / news_translator, etc.)
③Feedback-related (compare / ingest_manus / rule_curator)
④Testing & Libraries (test_lib / claude_client / quality_gate, etc.)





I didn't design all of this from the start.
It just grew one by one.

The moment I feel, 'It's a hassle to do this same process manually,' I first consult my secretary.
For simple tasks, I bounce ideas off it in this flow before handing them over to Automation/QA.

→ It grew naturally with the flow of scripts being born within a few days🚀




After bouncing ideas around and realizing, 'Ah, it would be better to create a script for this...,' this is how I do it.




⭐️Superpowers plugin and instructed my secretary to use it for implementation. I decided to handle things by having
“Implementation planning (brainstorming, etc.) → Secretary
Actual development (coding) → Development Department (Automation/QA Department)”
handle the respective tasks.

Using Superpowers for implementation is genuinely highly effective. Seriously.




Superpowers plugin is explained thoroughly in the article below, so if you are not familiar with it, please check it out👇




🔴 Traps discovered during operation⚠️

I have written only about the good things, but I will also honestly write about the traps I discovered after running the 5-department system for a month.



⚠️ Here are 3 common traps to fall into👇

💀Trap 1: Creating too many departments leads to chaos
　→ Creating departments on a whim will lead to immediate collapse. *On X, there have been reports of cases where '40 agents were built, but the system collapsed and was completely scrapped within a month'😭

💀

Trap 2: Context bloat (Context Rot)　→ If you keep talking to the secretary in the same session for a long time,

the token count increases and the secretary's judgment accuracy drops.

*Delegate heavy development work to separate agents, split chats, etc., to avoid task specialization—or rather, 'AI agent specialization.' Generalists have it tough, don't they? It's the same as with humans.💀

Trap 3: Getting used to the initial setup
　→ Slash commands like `/plugin` or `/company`, 　　and Markdown-centric operations require familiarity with the terminal.
*

Complete beginners will need 1-2 days of trial and error

All of these are things you can prevent if you know about them beforehand, so if you are just starting out, please keep them in the back of your mind💡




It works better to separate agents rather than departments. That's my gut feeling.



Chapter 5: Information Sharing Rules and Daily Reporting Flow — Systems That Become More Effective as the Organization Grows✅



As the number of departments increases, the importance of information sharing rules increases exponentially.

I've really felt this firsthand🔥



When I switch between AI agents to work, I sometimes end the day having worked in the editorial department or other departments.



What becomes important here is for the editorial department and each department to report their work status to the secretary.



Since AI agents do not share context (memory) with each other,



the secretary doesn't know what has been done or how far it has progressed, and when I try to continue from yesterday, the conversation with the secretary doesn't align.



To prevent this, it is necessary to establish daily reporting rules.




The image looks something like this.




🟠 Without rules, 5 departments will immediately descend into chaos

The moment it increased to 3 departments, communication errors occurred frequently.



Things like, "Wait, where did I write down this decision?" or "Whose inbox is the new task in?"

You might think the secretary has everything under control, but surprisingly, communication errors do happen.




Therefore, I have included these rules in the company itself's "CLAUDE.md" file.





In my case, I have also added a daily report submission rule to the CLAUDE.md of each department to ensure compliance with company rules.



Now, as for the specific rules, here is what we decided after bouncing ideas back and forth.

📗 Here is the information coordination rule (summary) finalized for 5-department operation👇
　✅Summary: Each department submits a daily report to the Secretary's Office upon completing work.
　✅Submission required regardless of work status (includes days with no work).
　✅Departments subject to reporting: Secretary's Office, Editorial Dept, SNS Ops Dept, Automation/QA, Shared
　✅Save location: secretary/inbox/YYYY-MM-DD-daily-report-[department].md (department suffix: secretary / editorial / sns-ops / automation-qa / shared)

I realized later that submitting a daily report when there is 'no work' requires creating a dedicated script.
This is because a report cannot be submitted while the agent is not running. This is a point to be careful about⚠️




I am writing down the flow for submission from each department to the secretary.





🤖 Here is how the secretary handles daily report management to ensure these rules are followed.

🤖Secretary's Office collection steps
Every morning (9:00), the morning roll call script is automatically launched via launchd
Automatically checks the submission status of the previous day's reports from each department
Records submission status, overdue tasks, and unreflected tasks in the morning roll call report (secretary/inbox/YYYY-MM-DD-morning-rollcall.md)
Notifies the owner via macOS notification if any issues are detected

🤖Secretary's Office response
Checks the morning roll call report and examines 'discovered/new tasks'
Adds to Notion '📝 Secretary's Office TODO' as needed
Adjusts priority and assigned department, then reflects changes

🤖Automatic check specifications
The morning roll call script determines the following.
Submitted ✅ Normal: File exists + body does not contain 'no work'
Submitted ⚠️ No work: File exists + body contains 'no work'
Not submitted ❌: File does not exist

🤖Morning roll call report summary line format:
Unreflected: A items / Overdue: B items / Not submitted: C departments / Total: W items
✅When in doubt, use `inbox/`: Don't aim for perfection, just throw it in for now
✅Date check required before file operation: Prevents AI-specific date errors



This is subtle, but it has the robustness to be extended without breaking.




🔵 The daily report flow is actually the most fun part

✅ This is the part I most wanted to write about in Part 2.

I've been running a mandatory departmental daily report operation for a month now, and
the experience of seeing Notion Todo items accumulate daily is quite fun👍



✅ The flow looks like this.

📘 Daily report flow running every morning (5 steps)👇
・Every morning, the secretary asks all departments, "Please provide your work report for today"
・Each department agent submits a daily report — The format is standardized
・Daily reports are saved to `secretary/inbox/YYYY-MM-DD-daily-report-[Department].md`
・The secretary reviews the daily reports → Automatically adds "Discoveries/New Tasks" to the Notion "📝 Secretary's Office TODO"



✅ The daily report format looks like this.

📙[Standard Daily Report Format]
## [Daily Report] YYYY-MM-DD | Department Name
### Completed Tasks - [x] Task Name
### Tasks in Progress - [ ] Task Name | Progress: XX%
### Discoveries/New Tasks - [NEW] Task Content | Priority: High/Normal/Low | Details: Background/Reason
### Issues/Blockers - Issue Content (if any)
### Next Steps - Next Action

📙[No-Work Daily Report Format]
## [Daily Report] YYYY-MM-DD | Department Name
### Completed Tasks - No work (Reason: Owner out of office / No assigned tasks, etc.)


Since the rules are clearly stated in Markdown, every agent can operate under the same standards✅




This flow offers an experience that could never be created with a single chat AI.

It's that feeling of realizing the system is recording progress on its own.



Every morning when I open Notion, what I did yesterday and what I need to do today are already listed.
I can check it from my smartphone too.



I've narrowed the status down to only those that are not yet started. There are quite a lot, aren't there?😅




🤖"How is this different from writing it manually?"🤖

I can hear that question, but this is completely different.



The fact that it piles up even though I didn't write it myself is huge. I just send one line saying, "I'll be receiving your work reports today too."




After that, the secretary and the agents for each department handle it automatically.

In terms of saving disposable time for side-hustlers, this is truly effective💪



And, it is simply fun to watch the Todo list pile up in Notion.




It feels similar to having your daily battle results automatically tallied in a game.

Since the visualization of results happens every day, my motivation for my side hustle naturally increases as well 🚀



Chapter 6: A Meaningful PDCA Loop — Scripts and AI Nurturing Each Other 🎯


Finally, here is the story of the mechanism that is the most effective.



A PDCA cycle that revolves around the triangle of Script × AI Agent × Rules.




🟢 Concrete Example of a Feedback Loop

I won't be able to convey this unless I write it concretely, so I will go with a real-world example.


📘 Here is the 6-step feedback loop!👇
①User realizes while working, 'Oh, it's faster to automate this'
→ Share with secretary
②Secretary brainstorms with user, requests Automation/QA
→ Scripting
③The relevant process is automatically executed from next time
→ AI agent can focus on other judgment tasks
④AI reviews the script's output
→ Provides feedback like 'I want this output to be a bit more like this'
⑤Feedback is promoted to writing style rules or quality standards
 → Recorded in `style-guide/` or `learned-rules.md`
⑥Rules are automatically injected during the next AI agent invocation
 → All agents benefit




Doing it just once is plain.

Running it every day creates a compounding effect🔥

My gut feeling is that 60% of the tasks I was doing manually a month ago have now been automated.



Here is the general idea.




🟣 AI compensates for what scripts are not good at

What is interesting about this design is that it creates a structure where scripts and AI compensate for each other's weaknesses.


✅ This structure of compensating for weaknesses can be expressed in these two lines👇
　📙 Scripts are fast, cheap, and have excellent reproducibility, but they are bad at contextual judgment
　🤖 AI is good at contextual judgment, but calling it every time adds up in cost

That is why the division of labor—'scripts for routine processing' and 'AI only for parts requiring judgment'—is decided automatically.




After operating it for a month, you develop an intuition for where to draw the line👀



🔵 A message that side-hustlers, more than anyone, should be running the PDCA cycle

The greatest enemy for a side-hustler is time.



In the limited disposable time between your main job, everything depends on how your systems can self-replicate.



What I realized this past month was that the moment the four gears of secretary x department x script x rules mesh together, there are times when your side-hustle system moves forward even without you lifting a finger🎯



Honestly, at first I thought, "Isn't it a bit much for a side-hustler to talk about systemization?"

But, you can't sustain a side-hustle if you don't systemize it,
"If the system grows, your side-hustle won't stop even on days when you're exhausted from your main job."


This is the conclusion I reached after one month of operation✅


Summary🎯

In short, these are the four things I wanted to write about across both parts.


📘 4 messages spanning both parts👇
✅Single-chat AIs do not become assets. The structure where conversations flow away and disappear is fatal for side-hustlers (Part 1)
✅A Claude Code secretary starts up in 15 minutes, and assets grow like an RPG in one month. This is the true nature of the feeling I wrote about in Part 1: 'I got back over 100 minutes of disposable time a day' (Part 1)
✅Turn tasks you repeat more than twice into departments. This is the next phase of the RPG, where your solo self becomes a party (Part 2)
✅Systems self-replicate through daily reports + Notion integration + feedback loops. The reason over 40 scripts are running is because of this trinity (Part 2)




This article is a field observation log of how I started with a secretary and realized it had become an organization.



Summary of this installment!



"The less time you have as a side-hustler, the more this route fits you." I've realized this after testing it for a month✅



In the first part, I posed this question.

💬"Where did I leave this conversation with the AI?"

Chat AI has its merits, but for continuous tasks...




Now that you've read the second part, I'd be happy if your answer has changed a little.

🏢Even if conversations flow away, the organization remains.
Your digital clones grow, become a party, and start running alongside your scripts.🚀

I would be truly happy if you could empathize with this.



If you are currently tired of single-chat AI, or if your tasks are scattered and your ideas are slipping away.



Just running `/company` on your way home today should change how you interact with AI starting tomorrow.


Just answer two questions. It really starts with just that💪



If you haven't read Part 1 yet, please check it out too👀



Thank you for reading until the end‼️



So, what kind of AI tool would you like to try next⁉️
Please let me know your thoughts in the comments‼️


If you found this article even a little bit helpful, clicking the 'Like' button would be a huge encouragement for me to write the next one‼️

いいなと思ったら応援しよう！