[GH-ISSUE #42] Thanks and Questions

kerem commented

2026-03-03 18:50:07 +03:00

Owner

Originally created by @jiangyixing on GitHub (Jan 19, 2026).
Original GitHub issue: https://github.com/OthmanAdi/planning-with-files/issues/42

Hello，Ahmad Othman Ammar Adi。Thank you very much for sharing. After reading about your project, I was indeed inspired, but I have a question about how you determined that this is the workflow of Manus？

Originally created by @jiangyixing on GitHub (Jan 19, 2026). Original GitHub issue: https://github.com/OthmanAdi/planning-with-files/issues/42 Hello，Ahmad Othman Ammar Adi。Thank you very much for sharing. After reading about your project, I was indeed inspired, but I have a question about how you determined that this is the workflow of Manus？

kerem closed this issue

2026-03-03 18:50:08 +03:00

kerem commented

2026-03-03 18:50:08 +03:00

Author

Owner

@OthmanAdi commented on GitHub (Jan 19, 2026):

hello friend. i went in detail on this topic on reddis ^_^. let me make a discussion post and share with you exafctly how i built it? sounds good? @jiangyixing

@OthmanAdi commented on GitHub (Jan 19, 2026): hello friend. i went in detail on this topic on reddis ^_^. let me make a discussion post and share with you exafctly how i built it? sounds good? @jiangyixing

kerem commented

2026-03-03 18:50:09 +03:00

Author

Owner

@jiangyixing commented on GitHub (Jan 19, 2026):

hello friend. i went in detail on this topic on reddis ^_^. let me make a discussion post and share with you exafctly how i built it? sounds good? @jiangyixing

Haha, that's great! If possible, I think it could also be placed on the homepage, so that more people can see it

@jiangyixing commented on GitHub (Jan 19, 2026): > hello friend. i went in detail on this topic on reddis ^_^. let me make a discussion post and share with you exafctly how i built it? sounds good? [@jiangyixing](https://github.com/jiangyixing) Haha, that's great! If possible, I think it could also be placed on the homepage, so that more people can see it

kerem commented

2026-03-03 18:50:09 +03:00

Author

Owner

@OthmanAdi commented on GitHub (Jan 22, 2026):

How I Determined This Is the Manus Workflow

Great question @jiangyixing! Here's the evidence from official sources:

Primary Source: Official Manus Blog

The official Manus blog article "Context Engineering for AI Agents: Lessons from Building Manus" describes their exact approach:

1. The todo.md Pattern (Direct Quote)

"When handling complex tasks, Manus tends to create a todo.md file—and update it step-by-step as the task progresses, checking off completed items. That's not just cute behavior—it's a deliberate mechanism to manipulate attention."

2. Why It Works (Direct Quote)

"A typical task in Manus requires around 50 tool calls on average. That's a long loop—and since Manus relies on LLMs for decision-making, it's vulnerable to drifting off-topic or forgetting earlier goals. By constantly rewriting the todo list, Manus is 'reciting its objectives into the end of the context.'"

3. File System as Memory (Direct Quote)

"We treat the file system as the ultimate context: unlimited in size, persistent by nature, and directly operable by the agent itself. The model learns to write to and read from files on demand—using the file system not just as storage, but as structured, externalized memory."

The Core Principle

Manus Concept	This Plugin's Implementation
todo.md for attention manipulation	task_plan.md with phases
File system as external memory	findings.md, progress.md
Re-read goals to prevent drift	PreToolUse hook reads plan
50+ tool calls problem	Same problem in Claude Code

Secondary Source: Lance Martin's Analysis

Lance Martin's deep dive confirmed:

Manus uses a 3-agent architecture: planner, knowledge manager, executors
Initial todo.md approach evolved to dedicated planner agent
~1/3 of actions were spent updating the todo list (token efficiency issue)

My Implementation

I took the core principles and adapted them for Claude Code:

task_plan.md = todo.md (attention manipulation)
findings.md = knowledge storage (file system as memory)
progress.md = session log (error tracking)
Hooks = automatic re-reading (attention refreshing)

This isn't a copy of Manus's code—it's an implementation of the same principles they discovered through building a production AI agent.

Additional Resources

Hope this explains the research behind the project! 🙏

@OthmanAdi commented on GitHub (Jan 22, 2026): ## How I Determined This Is the Manus Workflow Great question @jiangyixing! Here's the evidence from official sources: ### Primary Source: Official Manus Blog The official Manus blog article **["Context Engineering for AI Agents: Lessons from Building Manus"](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)** describes their exact approach: #### 1. The todo.md Pattern (Direct Quote) > "When handling complex tasks, Manus tends to create a **todo.md** file—and update it step-by-step as the task progresses, checking off completed items. That's not just cute behavior—it's a deliberate mechanism to manipulate attention." #### 2. Why It Works (Direct Quote) > "A typical task in Manus requires around **50 tool calls** on average. That's a long loop—and since Manus relies on LLMs for decision-making, it's vulnerable to **drifting off-topic or forgetting earlier goals**. By constantly rewriting the todo list, Manus is 'reciting its objectives into the end of the context.'" #### 3. File System as Memory (Direct Quote) > "We treat the **file system as the ultimate context**: unlimited in size, persistent by nature, and directly operable by the agent itself. The model learns to write to and read from files on demand—using the file system not just as storage, but as **structured, externalized memory**." ### The Core Principle | Manus Concept | This Plugin's Implementation | |--------------|------------------------------| | todo.md for attention manipulation | task_plan.md with phases | | File system as external memory | findings.md, progress.md | | Re-read goals to prevent drift | PreToolUse hook reads plan | | 50+ tool calls problem | Same problem in Claude Code | ### Secondary Source: Lance Martin's Analysis [Lance Martin's deep dive](https://rlancemartin.github.io/2025/10/15/manus/) confirmed: - Manus uses a **3-agent architecture**: planner, knowledge manager, executors - Initial todo.md approach evolved to dedicated planner agent - **~1/3 of actions** were spent updating the todo list (token efficiency issue) ### My Implementation I took the core principles and adapted them for Claude Code: 1. **task_plan.md** = todo.md (attention manipulation) 2. **findings.md** = knowledge storage (file system as memory) 3. **progress.md** = session log (error tracking) 4. **Hooks** = automatic re-reading (attention refreshing) This isn't a copy of Manus's code—it's an implementation of the same **principles** they discovered through building a production AI agent. ### Additional Resources - [Manus Blog](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus) - [Lance Martin Analysis](https://rlancemartin.github.io/2025/10/15/manus/) - [DEV.to Summary](https://dev.to/contextspace_/context-engineering-for-ai-agents-key-lessons-from-manus-3f83) Hope this explains the research behind the project! 🙏

kerem commented

2026-03-03 18:50:09 +03:00

Author

Owner

@jiangyixing commented on GitHub (Jan 23, 2026):

How I Determined This Is the Manus Workflow

Great question @jiangyixing! Here's the evidence from official sources:

Primary Source: Official Manus Blog

The official Manus blog article "Context Engineering for AI Agents: Lessons from Building Manus" describes their exact approach:

1. The todo.md Pattern (Direct Quote)

"When handling complex tasks, Manus tends to create a todo.md file—and update it step-by-step as the task progresses, checking off completed items. That's not just cute behavior—it's a deliberate mechanism to manipulate attention."

2. Why It Works (Direct Quote)

"A typical task in Manus requires around 50 tool calls on average. That's a long loop—and since Manus relies on LLMs for decision-making, it's vulnerable to drifting off-topic or forgetting earlier goals. By constantly rewriting the todo list, Manus is 'reciting its objectives into the end of the context.'"

3. File System as Memory (Direct Quote)

"We treat the file system as the ultimate context: unlimited in size, persistent by nature, and directly operable by the agent itself. The model learns to write to and read from files on demand—using the file system not just as storage, but as structured, externalized memory."

The Core Principle

Manus Concept This Plugin's Implementation
todo.md for attention manipulation task_plan.md with phases
File system as external memory findings.md, progress.md
Re-read goals to prevent drift PreToolUse hook reads plan
50+ tool calls problem Same problem in Claude Code

Secondary Source: Lance Martin's Analysis

Lance Martin's deep dive confirmed:

Manus uses a 3-agent architecture: planner, knowledge manager, executors

Initial todo.md approach evolved to dedicated planner agent

~1/3 of actions were spent updating the todo list (token efficiency issue)

My Implementation

I took the core principles and adapted them for Claude Code:

task_plan.md = todo.md (attention manipulation)

findings.md = knowledge storage (file system as memory)

progress.md = session log (error tracking)

Hooks = automatic re-reading (attention refreshing)

This isn't a copy of Manus's code—it's an implementation of the same principles they discovered through building a production AI agent.

Additional Resources

Manus Blog

Lance Martin Analysis

DEV.to Summary

Hope this explains the research behind the project! 🙏

That sounds fantastic! Thank you for your time

@jiangyixing commented on GitHub (Jan 23, 2026): > ## How I Determined This Is the Manus Workflow > Great question [@jiangyixing](https://github.com/jiangyixing)! Here's the evidence from official sources: > > ### Primary Source: Official Manus Blog > The official Manus blog article **["Context Engineering for AI Agents: Lessons from Building Manus"](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)** describes their exact approach: > > #### 1. The todo.md Pattern (Direct Quote) > > "When handling complex tasks, Manus tends to create a **todo.md** file—and update it step-by-step as the task progresses, checking off completed items. That's not just cute behavior—it's a deliberate mechanism to manipulate attention." > > #### 2. Why It Works (Direct Quote) > > "A typical task in Manus requires around **50 tool calls** on average. That's a long loop—and since Manus relies on LLMs for decision-making, it's vulnerable to **drifting off-topic or forgetting earlier goals**. By constantly rewriting the todo list, Manus is 'reciting its objectives into the end of the context.'" > > #### 3. File System as Memory (Direct Quote) > > "We treat the **file system as the ultimate context**: unlimited in size, persistent by nature, and directly operable by the agent itself. The model learns to write to and read from files on demand—using the file system not just as storage, but as **structured, externalized memory**." > > ### The Core Principle > Manus Concept This Plugin's Implementation > todo.md for attention manipulation task_plan.md with phases > File system as external memory findings.md, progress.md > Re-read goals to prevent drift PreToolUse hook reads plan > 50+ tool calls problem Same problem in Claude Code > ### Secondary Source: Lance Martin's Analysis > [Lance Martin's deep dive](https://rlancemartin.github.io/2025/10/15/manus/) confirmed: > > * Manus uses a **3-agent architecture**: planner, knowledge manager, executors > * Initial todo.md approach evolved to dedicated planner agent > * **~1/3 of actions** were spent updating the todo list (token efficiency issue) > > ### My Implementation > I took the core principles and adapted them for Claude Code: > > 1. **task_plan.md** = todo.md (attention manipulation) > 2. **findings.md** = knowledge storage (file system as memory) > 3. **progress.md** = session log (error tracking) > 4. **Hooks** = automatic re-reading (attention refreshing) > > This isn't a copy of Manus's code—it's an implementation of the same **principles** they discovered through building a production AI agent. > > ### Additional Resources > * [Manus Blog](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus) > * [Lance Martin Analysis](https://rlancemartin.github.io/2025/10/15/manus/) > * [DEV.to Summary](https://dev.to/contextspace_/context-engineering-for-ai-agents-key-lessons-from-manus-3f83) > > Hope this explains the research behind the project! 🙏 That sounds fantastic! Thank you for your time

Rows
Columns

[GH-ISSUE #42] Thanks and Questions #28

How I Determined This Is the Manus Workflow

Primary Source: Official Manus Blog

1. The todo.md Pattern (Direct Quote)

2. Why It Works (Direct Quote)

3. File System as Memory (Direct Quote)

The Core Principle

Secondary Source: Lance Martin's Analysis

My Implementation

Additional Resources

How I Determined This Is the Manus Workflow

Primary Source: Official Manus Blog

1. The todo.md Pattern (Direct Quote)

2. Why It Works (Direct Quote)

3. File System as Memory (Direct Quote)

The Core Principle

Secondary Source: Lance Martin's Analysis

My Implementation

Additional Resources