Building a Task Management AI Based on HermesAgent Philosophy Turned Task Processing into 'Brain-Dead Wanko Soba'

I was too lazy to do it myself, so I let the AI do it. The AI breaks down tasks like wanko soba and throws them into my mouth. I've become a robot processing them on autopilot. I wonder what it means to be alive.

When I built a task management AI based on the HermesAgent philosophy, task processing became "brain-dead wanko soba" style 🔥

Inspired by the philosophy of Nous Research's HermesAgent, I built a "system that gets smarter every time a task is completed" using Claude Code + Markdown + shell hooks.

As a result, task processing became like "brain-dead wanko soba."

You just throw things in roughly using natural language. The AI senses the context and places it, the AI accumulates knowledge upon completion, and the AI flows the next thing to do. I can just stop thinking.

The era of "thinking to manage tasks" is over.

*This isn't a dedicated app. It's a general-purpose mechanism running on Claude Code.

━━━━━━━━━━━━━━━━━━━━━━

◆ First, the conclusion ── What is "brain-dead wanko soba" style?

Do you know "wanko soba"? It's a specialty of Morioka where a server continuously tosses noodles into your bowl, and the eater is swept along without being able to stop.

The task management system I built on Claude Code realizes this same experience for task processing:

Me: Roughly throw a task to Claude

Example: "Add a task to finalize Agenda A for Project A"

↓

Claude: Senses from past history/records → Automatically places it in the appropriate folder

Example: Writes to work/project-a/agenda-1/tasks/{id}-topic-a.md

↓

Claude: "Added task #67 to ... [Is this wrong?]"

↓

Me: Ignore it or throw the next thing like "I'll do this"

↓

Claude: Flows the next operation

"Then, shall we organize the draft A/B/C plans for Agenda A in #67?"

↓

Me: Just reply "Do it" or "OK"

↓

Claude: Executes work → Reports completion

"Promoted knowledge to Epic, Roadmap ☐→☑, closed the task"

↓

[Repeat, never stops, no need to think]

↓

Tasks are consumed one after another in a brain-dead state ✨

→ This is "brain-dead wanko soba task processing."

→ Fits the ADHD brain perfectly.

Implementation is simple: Claude Code + Markdown files + just 2 shell hooks. There is no dedicated app (I plan to make one in the future, but currently Claude Code is the UI).

And the core that realizes this is another power word:

【AI that gets smarter every time it completes a task】.

Ordinary task management tools just archive completed tasks and that's it. Nothing is born from there.

But the system I built is different:

Task completion → AI executes 4 steps in the background:

① Feedback ☐→☑ to the roadmap

② Automatically promote "decisions," "premises," and "learnings" gained in that task to the project note

③ Update the Story's Acceptance Criteria

④ Close the task

→ Individual tasks are archived after 14 days

→ But the knowledge accumulated in the project note (_epic.md) remains

→ When handling the next task, the AI acts based on that knowledge

In short, the AI gets smarter every time it completes a task ⭐

━━━━━━━━━━━━━━━━━━━━━━

◆ How this AI works ── 3-layer knowledge management + automatic knowledge promotion

◇ Mechanism 1: 3-layer knowledge base

My AI holds knowledge in three layers.

[Global] CLAUDE.md

↓ AI's principles of action / common rules for all projects

[Domain] {Hierarchy}/_README.md

↓ Domain-specific terminology, stakeholders, operation rules

[Project] {Project}/_epic.md

↓ Everything about the project (Why/What/How/When/Status/Knowledge/Retrospective)

[Task] {Project}/tasks/{id}-xxx.md

↑ Individual work files

When handling a task, the AI follows the parent hierarchy and reads everything.

Meaning:

• When looking at a task, it reads the _epic.md of that project
• When looking at _epic.md, it reads the _README.md of that domain
• When looking at _README.md, it reads the global CLAUDE.md

→ Even in a new AI session, past context is perfectly inherited.

◇ Mechanism 2: "Automatic Knowledge Promotion" upon task completion ⭐

This is the core.

Task files have a section called "📤 Epic Promotion Memo." Decisions, premises, and learnings that should remain in the project are stored here.

When setting a task to "complete," the AI does this in the background:

Reads the "📤 Epic Promotion Memo" in the task

↓

Automatically transcribes to the appropriate Epic section:

├ Specification finalized → Add a row to the "✅ Finalized Specs" table

├ Strategy/Judgment → "🧠 Strategy by Topic" section

├ Chronological facts → "📅 Timeline" table

└ Business terms → "📖 Glossary"

↓

Close task (epic_promoted: true, status: completed)

What happens because of this:

Individual task file

↓ 14 days pass

↓ Move to archive (disappears)

But the "Decision Log" and "Finalized Specs" inside _epic.md remain ⭐

↓

When handling the next task, the AI refers to this knowledge base

↓

Can provide answers based on past decisions

↓

Every time a task is completed, the project note (= AI's knowledge base) grows

↓

The AI gets smarter

This is the mechanism that turned "Completion = Discard" into "Completion = Learning."

━━━━━━━━━━━━━━━━━━━━━━

◆ Happiness born from this

◇ Happiness ① Knowledge accumulates for each domain

Project A's _epic.md:

...

Every day, knowledge accumulates. By the time 3 months pass, a manual for that project is automatically completed.

Even if a new member joins, just giving them _epic.md is enough. The time humans spend writing documentation vanishes.

◇ Happiness ② Completion is no longer the destruction of knowledge

Normal task management:

Task completion → Archive → Knowledge lost → 1 month later "How did we decide this?"

My AI:

Task completion → Auto-promote knowledge to project note → Knowledge persistence → Referrable even 1 year later ⭐

I can now answer "How did we decide that back then?" in 3 seconds.

◇ Happiness ③ An AI that gets smarter every time it completes a task was born

This is seriously amazing.

Day 1: Task #67 completed

AI: Records "Agenda A finalized with Plan A" in _epic.md

Day 5: I say "I want this kind of spec change related to Agenda A"

AI: "If we assume Plan A, it will be like this" (Answers based on Day 1 knowledge)

Day 30: Task #82 completed

AI: Records "Added category column to history table"

Day 45: Same pattern occurs in another project

AI: "In the previous Project A, we handled it by adding a category column"

(Recycles Day 30 knowledge)

→ AI's knowledge base expands daily

→ Never fall into the same hole twice

→ The more projects it crosses, the more the AI gains "industry common sense" ⭐

This is the same philosophy as the Self-improving loop of Nous Research's HermesAgent. I've reached the same place as the cutting edge of the AI industry (I don't think it's a coincidence).

◇ Happiness ④ The "Sensing Butler" grows with experience

Day 1: "Buy buttons"

AI: "What kind of buttons?" (Unknown)

Me: "For antique collecting"

↓ Recorded in AI's chart

Day 30: "Buy buttons"

AI: "Is antique style okay? Shop XX?" (Confidence 0.7)

Me: "OK"

↓ Confidence +0.1

Day 90: "Buy buttons"

AI: [Silently generates a cup based on Shop XX antique buttons] (Confidence 0.95)

Me: Even if I say nothing, the appropriate answer comes out ✨

The AI understands me. The "Sensing Butler" grows with experience.

◇ Happiness ⑤ Complete handover to other devices/AI sessions

Even if I start a different AI session on another PC, the AI has all past knowledge.

New AI: [Startup] → Reads CLAUDE.md / _README.md / _epic.md hierarchically

↓

Me: "Maybe I'll do Project A"

↓

New AI: "In the middle of M1 requirement definition, waiting for manager judgment on Agenda 1. Yesterday, Agenda A was finalized with Plan A. Today's 'Now' is Story X. What will you do?"

Me: "How do you know that?"

New AI: "I read _epic.md" 😎

I haven't explained anything. But the new AI completely understands the context.

━━━━━━━━━━━━━━━━━━━━━━

◆ Real-life scenes

◇ Scene: Morning task addition

Me: "Add the button buying thing"

↓

AI: Senses "Button = Antique, Shop XX" from past charts

↓ Automatically generates task file in Hobbies/Button Collecting/tasks/

↓

AI: "Sensed: Is Shop XX antique buttons okay? Task registered [Wrong]"

Me: Ignore (It's correct, so I do nothing)

What I did: Just said one sentence 🤯

◇ Scene: New project occurrence

Me: "I want to do a Musou game and 3DGS display with Unreal Engine's UEFN!"

↓

AI: Senses "UnrealEngine," "UEFN," "3DGS"

↓ No match in existing scope

↓ But can sense context → Automatically generates new scope

↓

AI: "Created new scope:

📂 Hobbies/UnrealEngine/UEFN-Musou-x-3DGS/

Tasks also placed. Shall we do an Epic-making dialogue tomorrow morning?"

Me: I did nothing. The AI created folders and placed them on its own.

Folder structures are made by the AI. No need to decide the structure from the start.

◇ Scene: The magic of task completion

Me: "#67 complete!"

↓

[AI executes 4 steps in the background]

① Update roadmap (□→■)

② Read "📤 Epic Promotion Memo" → Transcribe to _epic.md:

③ Update Story's AC

④ Close task

↓

AI: "#67 completion processing finished. Added one line of knowledge to Epic, the project note has grown fatter 📚"

Me: "Thanks" (Actually did nothing)

Every time it completes, the AI is getting smarter. It's accumulating knowledge on my behalf.

◇ Scene: Me one month later

Me: "What was happening with that project?"

↓

AI: Immediately reads _epic.md → Presents summary

↓

AI: "📋 Project Summary:

🚦 Now: Story X

🚧 Blocker: Waiting for manager judgment on Agenda 1 (since 3 days ago)

📌 Recent Decisions:

• 6/18: Agenda A finalized with Plan A (Manager approved)
• 6/18: Pop-up finalized with DB method
• 6/20: Agreed on adding category column to history table 📅 Next Milestone: 6/22 External partner consultation Overall Progress: 25%"

Me: Full context recovery in 5 seconds ⚡

What would normally take 30 minutes of "remembering work" ends in 5 seconds. All information is stored by the AI.

━━━━━━━━━━━━━━━━━━━━━━

◆ Benefits Galore ── What changed with this?

✅ Task processing is brain-dead wanko soba style ⭐ No more thinking, it just flows

✅ Completion = Knowledge persistence (usually completion = knowledge destruction)

✅ AI gets smarter every time it completes ⭐ Implementation of HermesAgent philosophy

✅ Knowledge base grows for each domain (Knowledge accumulates for Project A, Project B... respectively)

✅ "What happened with that case?" comes out in 3 seconds (Project note becomes a manual)

✅ Complete handover to new AI sessions (Just follow the 3 layers)

✅ Sensing Butler grows with experience (Day 1 confidence 0.3 → Day 90 confidence 0.95)

✅ Morning task organization: 30 mins → 3 mins (AI senses it for me)

✅ No Notion / Jira subscription needed (Saves tens of thousands of yen per year)

✅ ADHD-like confusion drastically reduced (No need to organize because structure grows on its own)

✅ AI that learns from completed tasks (Same philosophy as Nous Research's Self-improving loop)

━━━━━━━━━━━━━━━━━━━━━━

◆ Behind the scenes ── How was it implemented?

Surprisingly simple.

Source of Truth: Task files (Markdown + YAML frontmatter)

Derivatives: Dashboard (Automatically generated by machine)

Implementation:

~/.claude/bin/

├ regen-tasks-md # Walk folder hierarchy and generate dashboard

└ regen-epic-md # Automatically update Task list for Stories in Epic

Trigger:

PostToolUse hook (a feature of Claude Code)

The above scripts run automatically upon file editing.

No need for complex systems like Notion / Airtable / Jira.

OS file system + Markdown + hook = This is enough to work.

The fact that task management tools humanity spent 30 years building are surpassed by a few hundred lines of Markdown.

━━━━━━━━━━━━━━━━━━━━━━

◆ What is it based on? ── The culmination of 30 years

I didn't make this on a whim. It's the result of integrating the essence of 10 types of task management legends:

• Jira → Epic-Story-Task hierarchy + Acceptance Criteria
• Kanban → WIP limits, Now-Next-Later
• Scrum → Definition of Done + Retrospective
• GTD (David Allen) → Capture-Clarify-Organize-Reflect-Engage
• OKR → Objective + Measurable Key Results
• Linear → Triage Inbox (INBOX)
• Notion / Roam → Templates + Bi-directional links
• Bullet Journal → Migration (Archive)
• ADR (Architecture Decision Records) → Decision Log
• DDD (Domain Driven Design) → Bounded Context + Glossary
• HermesAgent (Nous Research) → Self-improving loop ⭐

I integrated 30 years of human knowledge into a form that AI can handle.

It's not "reinventing the wheel," it's the "culmination of the wheel."

━━━━━━━━━━━━━━━━━━━━━━

◆ Future ── Evolving this into a task app for ADHD

Actually, this is the design prototype for "World Milestone," a task app for ADHD that I'm developing personally.

Future:

• Appify with Rails 8.1
• Full implementation of "Sensing Butler"
• Serving micro-tasks with "short video grammar"
• Full implementation of "AI that gets smarter every time it completes"
• Planned for public release for individuals

The mechanism I'm sharing on X today is a complete design blueprint for that app.

When finished, it should definitely resonate with people with ADHD / people who waste time organizing Notion / people running multiple projects in parallel.

━━━━━━━━━━━━━━━━━━━━━━

◆ In Closing ── I want to create a new standard for task management

I'm serious, I can't go back to normal task management tools anymore.

Not "managing tasks," but "accumulating knowledge in AI."

Not "finishing upon completion," but "getting smarter every time it completes."

Not "thinking and organizing with your head," but "brain-dead wanko soba task processing."

I believe this is the new standard for task management.

Please remember these three power words:

🧠 AI that gets smarter every time it completes

🍜 Brain-dead wanko soba task processing

🤖 HermesAgent-inspired task management

It's a system from the future.

━━━━━━━━━━━━━━━━━━━━━━

Design details, templates, and implementation scripts are all planned for release.

Depending on the reaction, I'll put them on Github or summarize them in a note.

Please reply with your reactions/questions 👀

RTs / Likes are very encouraging 🙏

#TaskManagement #AI #ADHD #Productivity #NotionAlternative #JiraAlternative #SelfImprovingAI #ClaudeCode #HermesAgent