This article shares my usual learning, reasoning and coding toolbox. Everything listed here are what I personally find most useful, or even something I designed/developed for myself(and for everyone, in the future when it’s more like a product).
My workflow was built around a language-model centered principle to maximize their efficiency and minimize any flaws and complexities they brought along. The foundation is to gather and extract most valuable information that I gained from language model outputs and sort them into retrivable, structured knowledges, and indexed them for future uses.
1. Intelligence Layer
ChatGPT
Primary learning intelligence for most study tasks, including concept explanation, knowledge restructuring, proof clarification, and prompt-sensitive reasoning. Requires strict prompting and personal instruction constraints to avoid meta-formatting and output drift.
Claude
Secondary validation intelligence. Mainly used for review, verification, and checking whether a task has been completed correctly and cleanly.
Gemini
Fallback intelligence for stable, non-trending, objective domains such as mathematics, statistics, and machine learning. Best used under constrained prompting, while noting that excessive constraints may reduce output quality.
⸻
2. AGN 2.0
AGN is my own agentic production infrastructure. It has been evolved to another major update from 1.0 to 2.0 recently.
In DevMTRX section of my own website, there are a dedicated webpage demonstrating a series of design and decisions I made during the construction of AGN’s prototype.
Here I’ll only be introducing the roles of three agents that have long-term memory and communicate directly with me in the network.
Remote Assistant
Admin’s remote assistant outside the local system. Communicates with Admin through Telegram, can read and write within authorized file scopes, and can awaken AGN1.0 to perform research or other delegated tasks on demand.
Central Execution Unit
The real-time intelligence operating inside the active working platform. Uses tools, skills, and system permissions to help Admin reach specific outcomes under direct supervision.
The CEU is trusted because it operates under Admin’s real-time oversight. It has high-level system authorization and is optimized for controlled execution rather than autonomy.
Current model priority:
1. GPT-5.4 Codex
2. Claude
3. Cursor
4. Antigravity
When Codex is active, it becomes the CEU by default.
AGN1.0
The original AGN before 2.0 update. A sealed subsystem within AGN2.0 for semi-autonomous research and coding tasks with clear expected outcomes and execution instructions.
AGN1.0 consists of three non-Admin roles: • Coordinator • Executor • Reviewer
AGN is currently under construction, especially focused on the internet security and system transparency side, aiming to give human admin full control over a complex system’s behaviours, and being friendly for everyone regardless of your computer science background. Before achieveing that, AGN will be polished consistently.
I’m looking forward to share more details about the progress in the future, and make it open-source to everyone eventually. So stay tuned for more updates!
Resources
IDEs • PyCharm • VS Code • RStudio • Xcode
Infrastructure • OpenClaw
Available Models
Flagship Models • Claude Opus 4.6 • GPT-5.4 Codex High / Extra High • Gemini 3.1 Pro
Efficient Models • GPT-5.4 Codex Low / Mid • Claude Sonnet 4.6
Economical Models • DeepSeek v3.2 • Codex 5.1 Mini • Claude Haiku 4.6 • Gemini 3.1 Flash • Qwen / Qwen3.5-35B-A3B local model
⸻
3. Knowledge Management
XMind
Used for quickly sorting the hierarchy and structural relationships of knowledge.
Cognodex
Mindmap-like academic resource management system for storing and retrieving structured learning materials.
Obsidian
Primary note-taking system. Focused on atomic notes and graph-based interconnection between knowledge units.
NotebookLM
Source-constrained Gemini environment for overview building and question answering based on selected materials.
Notability
Used with Apple Pencil for handwriting, annotation, and diagram-based reinforcement.
Bear
Markdown capture tool for fast note intake. Later transferred into Obsidian and rewritten into atomic note format.
Marktap
Used for instant inspirations, temporary thoughts, and lightweight daily task capture.
PowerPoint
Used when many figures need to be inserted and organized in a single visual file.
⸻
- Utilities
Norkta
PDF processing utility.
Inserable
Hugo blog integrity management and Markdown editing utility.
Texmorph
Text formatting and repair utility for Markdown, LaTeX, and code.
Rhylm
White-noise utility for masking distractions during learning.
Backtrack
Recursive task and habit tracking tool for monitoring daily progress.
I’m constantly updating my toolbox as I move forwards with my learnings. Finding the best stuff for yourself takes time and efforts to adapt, understand and merge it into a way that works for you. What’s more important is to build a pipeline within tools you use.
For example, you may write on Notability and then let your preferred multimodality language model to extract your writings and formatting them into markdown and paste them into Bear notes, then you connect them with links and eventually put them into Obsidian. The process is accelerated because language models will do the repetitive work for you, so you only need to focus on the thinking side.
Think systematically is the most useful skill I’ve developed over the past few years. Think not in a way of how to use something, but how something’s could be utilized and apmlifies with consideration of their input and output mechanism.
In this way, everything will become “modular” for me, just like modular synthesizers could be splitted into different modules and reassembled into something completely new, while traditional synths won’t be able to reach this kind of flexibility due to I/O constraients. Becoming modular means you will be updating your system with compatible new stuff, treat your new modules like a blackbox first for everything else in the system, focus only on the input/output compatability before and after, so your system evolves with you.
Seems I’ve talked too much about a bit off-topic stuff. Appreciate your reading, and I’ll see you soon on next one!