Signal to BuildSIGNAL TO BUILD中文Owner sign in ↗
← Knowledge Notes
Personal Growth ／ Knowledge note · English

newtype Knowledge Management System: The Evolution of Obsidian, NotebookLM, and Life OS

Understand the evolution of the newtype knowledge system through long-term notes, source-based Q&A, and Agent execution; front-load the import boundaries of Obsidian and NotebookLM and the reading entry points, while preserving the original architecture, tools, and historical viewpoints.

Source collection：Newtype · Published here：2026-09-26

Personal Knowledge ManagementLife OSSecond Brain

First answer: What do Obsidian, NotebookLM, and the Agent each do in a knowledge system?

This article is an evolution record of a personal knowledge workflow. First distinguish three tasks: long-term storage and revision of one’s own notes; asking questions around selected materials and verifying sources; and having the Agent perform bounded organizing work. The key to connecting the three is data sources, update methods, and result verification—not installing the largest number of tools at once. The old draft’s line, “put your own things in Obsidian and other people’s things in NotebookLM,” is the author’s usage preference; who wrote the material alone cannot determine whether you have the right to upload or share it.

Supplementary source verification: 2026-10-25. Obsidian officially defines a vault as a local folder; this addresses the storage location, but “local” alone cannot imply that all plugin and model requests execute locally. Google’s current sources help distinguish imported copies from automatically synced sources: webpages import only the text of that page, excluding images, embedded videos, and subpages; public video imports use subtitle text. If you want to analyze visuals or an entire website’s content, you should separately check whether you have actually obtained those materials.

Official explanation:Obsidian vault,Google source types and import boundaries. Different sources have different update mechanisms; do not treat a successful import as bidirectional synchronization of an entire library. First use one document for which use is permitted to check whether answers after changes can cite the current version.

Choose the reading entry point by task

First verify once that it “can be found, asked accurately, and changed back”

This site recommends: choose only one public document or one document approved for processing, preserve the original source and date, organize one topic note with citations, and then ask one question the original text can answer and one question it cannot answer. Separately check the citation location, whether information is omitted, and whether insufficiency of materials is clearly stated; when having the Agent organize things, first preserve the pre-modification version, then check accidental deletion and recovery methods. This small exercise is meant to expose process gaps, not serve as a product comparison or efficiency benchmark of already running products.

Below, the “2026 final state,” “current recommendation,” model names, networking plans, prices, and installation commands belong to notes from the corresponding period and do not mean the system has stopped evolving or that every item has been reverified as a current recommendation. When you need to hand organizing tasks to an Agent, refer to Agent acceptance criteria; when you need to arrange a longer learning path, refer to Karpathy learning map.

1,085 posts, 26 months, 5 generations of architecture. From the AnythingLLM “slaughterhouse” to newtype OS “multi-agent orchestration.”

One sentence summarizing the 2026 final state:Put your own things in Obsidian, put other people’s things in NotebookLM, and Agent scheduling goes through newtype OS / Claude Code / Codex CLI.

0. Reading Route

If you do not want to read everything:

  • Entrepreneurs/independent developers → read §3 (three-layer architecture) + §5 (newtype OS) + §8 (productization)
  • General users/beginners → read §1.4 (ultimate conclusion) + §6.5 (NotebookLM) + §9.2 (Gemini subscription)
  • Learners → read §1 (core mental model) + §10 (Karpathy LLM Wiki) + §11 (Conversational Learning)
  • Want to see the evolution curve → read §2 (5 generations of architecture)

1. Core Mental Model / Lao Huang’s Knowledge Philosophy

1.1 Starting Point (Post 147, 2024-05 / First Systematic Exposition)

In Post 147 (2024-05-04, the earliest seminal long post in the knowledge-base field), Lao Huang systematically organized this for the first time:

  1. Information overload makes it impossible to digest everything one by one in detail → AI assistance helps us quickly and roughly grasp things through summarization and distillation.
  2. The human brain is not suited for memorizing things; it is suited for creative work → A “Second Brain” is needed for storage, while AI performs semantic retrieval (keyword retrieval + semantic retrieval).
  3. Taking notes is preprocessing information. The purpose of taking notes is to reinforce one’s future self → Information must be filtered layer by layer, and the selected content worth recording should be recorded in one’s own words rather than copied from other people’s expressions.

Why: These three points are the immovable anchors of Lao Huang’s two years of knowledge-base iteration. All later tool changes (AnythingLLM → Cursor → Claude Code → OpenCode) served these three points; the essence did not change.

1.2 Slaughterhouse + Second Brain (2024)

Two subsystems:

  • External information processing / slaughterhouse: AnythingLLM + a large model. Throw in information encountered in daily life and let the LLM, the “knife,” dismember it so that the “whole-body structure” and “valuable parts” can be seen clearly.
  • Note generation / Second Brain: Obsidian + various plugins. Manually filter information layer by layer, select what is most worth recording, and write it down in your own words.

Core ritual: It must pass through your own mind and be typed by your own hand. Otherwise it definitely will not become your own thing.

1.3 PAFP Four Folders + Three-Layer Linking

PAFP (Projects / Areas / Files / Personal) was Lao Huang’s early Obsidian directory structure.

Three-layer linking (Post 171, 2024-05, citing Musk’s OpenESG sharing):

  • Folders = first-layer linking
  • Tags = second-layer linking, cross-folder logic
  • Internal links / wikilinks = third-layer linking, knowledge graph

Musk’s words: “Memory is key to biological survival and an extremely precious resource, so the brain always forgets; this is a protection mechanism. If you want the brain to remember, you must establish relevance.”

Obsidian’s three-layer linking = establishing relevance for the brain = reducing forgetting. This is the fundamental reason Lao Huang has always chosen Obsidian rather than Notion.

1.4 Funnel Directory (Post 857, 2025-02 / Current Directory Structure)

By 2025, PAFP had converged into funnel logic:

  • Input = various articles, papers, and so on that one wants to examine closely (collection layer)
  • Processing = deposits organized by domain (digestion layer)
  • Output = final output; all video scripts are stored here (creation layer)

Why: PAFP is suited to static storage; the funnel is suited to an assembly line—information flows from Input to Processing to Output with a clear direction.

1.5 Ultimate Conclusion (Post 2201, 2026-01)

After two years of iteration, on 2026-01-03 Lao Huang gave the currently most stable conclusion:

The ultimate solution for a personal knowledge base:

  1. Put your own things in Obsidian.
  2. Put other people’s things in NotebookLM.

Why each:

  • Your own things: must be local-first; one cannot trust any company. Markdown documents are extremely AI-friendly and can be processed in a highly customized way with Obsidian AI plugins or Claude Code / OpenCode.
  • Other people’s things: no need to consider privacy; pursue performance. NotebookLM is the most powerful tool at present (multimodal, podcast generation, Source Grounding, Gemini invocation, mobile photo upload).

1.6 Three Immovable Anchors of Learning (Post 857, 2025-02)

Lao Huang’s answer when asked “how to arrange time and tools” (tools are secondary; the most critical things are these three):

  1. Vitality: the source drive; the level of vitality is the greatest difference between people. My output of vitality is creation and pioneering.
  2. Physical energy: Many people underestimate physical energy, but physical energy determines execution. Exercise every day, eat scientifically, and sleep sufficiently.
  3. No forcing: “Self-discipline determines freedom” is harmful—when “self-discipline” is involved, you are already not free, with a great deal of internal friction. So no self-discipline is needed; updating videos weekly and posting every day happen naturally.

1.7 Notes = Context in the AI Era (Post 1034, 2025-03 / Post 1075, 2025-04)

“In the AI era, notes are the Context between you and AI. If you want AI to help you, you have to let it understand you and give it enough ‘context.’” — 2025-03-26

Why: This is the underlying belief behind Lao Huang’s Prompt House, Obsidian MCP, and newtype Profile. In 2025-07, when answering a Star Friend’s question about “how to choose a product track,” he replied:Context Matters! This is the same underlying carrier for all his products.

1.8 Markdown Is Infrastructure in the AI Era (2025-03 post 998, citing Karpathy)

"Today, 99% of content is still created for humans rather than LLMs. Starting in 2025, all documents should be created in Markdown format."

How to apply: This is also why Lao Huang converted all his materials into md notes. PDF/PPT/docx files must first be converted to Markdown (using MinerU / Marker / Mathpix / Microsoft markitdown).

1.9 Learning by Doing Is Intuition Training (2025-10 post 1888)

"The doing in learning by doing is a form of intuition training."

Why this matters: Lao Huang opposes "learning by reading." Starting in 2023, when he self-funded front-end and back-end freelancers to build products (Chatbot + RAG + online search), through full-stack development with Codex in 2026, all his understanding of AI has come from hands-on practice.

2. Five Generations of Architectural Evolution (2024-02 → 2026-05)

Generation 1: AnythingLLM "Slaughterhouse" (2024 Q1 to Q2)

Core tool stack:

  • Ollama (running open-source models locally) / LM Studio (MLX support for Mac M-series)
  • AnythingLLM (knowledge base frontend, integrating all features)
  • Obsidian + Copilot/Text Generator (notes + AI plugins)
  • Open WebUI (#file document references)
  • QAnything (NetEase, connected to official-account customer service)
  • MaxKB (enterprise internal training)

Why:

  • AnythingLLM "integrates everything into it as the sole frontend" (assessment in 2024-03-30 post 44)
  • Workspace isolation: personal (local novels/e-books) + OpenAI (API)
  • Citations for sources (also available with local large models)
  • Drawbacks: Agent performance was rudimentary and often got stuck in loops (2024-04 post 111)

Pitfalls:

  • The Llama series had poor Chinese capabilities, so he switched to qwen-1.5 / qwen2
  • Ollama models default to downloading to the C drive → set the OLLAMA_MODELS environment variable
  • GPT-4 API errors on AnythingLLM → check the baseURL
  • Docker on macOS M3 had an "incompatible chip" issue (2024-04 post 127), and WSL was also painful

Generation 2: Cursor + Obsidian Knowledge Base (2024 Q3 to Q4)

Key turning point:

In 2024-08 (post 483), Lao Huang discovered:Open the Obsidian Vault with Cursor, @ a folder to ask questions, and the AI knowledge base is done!

"Cursor and Obsidian are a perfect match; it’s like a super AI plugin." (2024-09 post 487)

Core combination:

  • v0 (UI) → Claude (framework) → Cursor (details + debugging) (2024-09 post 488, the trio)
  • Obsidian + Web Clipper (replacing Omnivore, 2024-11)
  • Mac mini M4 + Enchanted (lightweight LAN AI server, 2024-12 post 677)

Core tools:

  • v0 (UI design, used in 2024-09 post 481 to build the newtype.pro Blog)
  • Cursor + Claude 3.5 Sonnet (main battlefield)
  • Three major Obsidian plugins: Copilot / Text Generator / Web Clipper
  • ComfyUI + Flux + LoRA (images)
  • Marker / MinerU / Mathpix / markitdown(PDF→MD)

Pitfalls:

  • One GraphRAG Q&A cost $11 (2024-07 post 360) → switched to local qwen2 + nomic, but had to heavily modify openai_embeddings_llm.py
  • ChatGPT access in China was unstable, with frequent refresh verification (2024-06 post 319) → moved to API + third-party software
  • Anthropic Claude accounts were frequently banned in China → switched to Gemini

Key migration: Omnivore shut down in 2024-11 → migrated entirely to Obsidian Web Clipper.

Generation 3: Cursor + Obsidian + Deep Research Trio Learning Method (2025 Q1 to Q2)

Core methodology (2025-02 post 854 / classic)

"A learning method I recently developed: Obsidian + Deep Research + Cursor:"

  1. Use Deep Research to complete an analysis of a topic (for example, an analysis of DeepSeek-R1’s technological innovations)
  2. Paste the content into Obsidian and open it with Cursor
  3. Use Cursor’s Composer mode to first tag all knowledge points with[[]] backlinks
  4. Have Cursor create blank md documents for every annotated knowledge point
  5. Fill in the documents, explain the knowledge points, and make them as accessible and easy to understand as possible

Final result: one topic analysis + all knowledge points + knowledge graph + an explanation for every knowledge point.

Cursor + .cursorrules knowledge base (2025-04 post 1154)

Explicitly require every answer to query all documents, specify how to use the 4 MCPs, and when to search online →Cursor becomes the most usable knowledge base software.

Key infrastructure leap:

  • Milvus + MCP local vector database (2025-04 post 1083) → separate the database from the model
  • Cline self-built MCP: Obsidian MCP + Replicate Flux MCP (2025-03 post 962)
  • mcp.so one-stop MCP store (2025-03 post 968 / made by idoubi)
  • 21st.dev Magic (2025-03 post 965, the first commercialized MCP)

Main model switches:

  • Primary: Grok 3 + Gemini 2.0
  • Auxiliary: ChatWise + OpenRouter + Ollama (local)
  • Primary search tool: Perplexity
  • Programming: Cursor + Claude Sonnet 3.7
  • Knowledge base: Obsidian + Copilot

Pitfalls:

  • When Manus came out, Sequential Thinking + Tavily simulated Deep Research, with one Q&A costing $1 (2025-03 post 1011)
  • Cursor network issues → disable HTTP2 (2025-05 post 1327)
  • An overly long context in the same chat window has an impact → open a new window when solving independent problems (2025-05 post 1201)

Generation 4: Claude Code AI OS (2025 Q3 to Q4)

Core belief shift (2025-07 post 1472)

"Unsubscribe from Cursor. It’s not just the $20 saved."

Afterward, Lao Huang officially shifted to Claude Code (subscribing to the $100 tier, with Opus accounting for 10% and Sonnet 4 used for the rest). Reasons:

  • Sub-agents (2025-08 post 1572; configure different models—Haiku for simple tasks, Opus for complex reasoning)
  • Plan Mode (Opus planning + Sonnet execution, balanced)
  • Custom commands (Slash Command) + Multi-directory (2025-09 post 1759)
  • CLAUDE.md > 40K prompt splitting (2025-08 post 1648)
  • /context View consumption (2025-08 post 1658)
  • /hooks configure sound-effect reminders (2025-08 post 1652)

Skills Are the Key (2025-10 post 1899 / classic)

Lao Huang made his first Skill: Super Analyst — 12 analysis frameworks (first principles / impossible trinity / SWOT / ...) stored in Prompt House; after triggering the Skill, detailed Prompts are pulled via MCP.

Afterward, Super Writer(seven-step SOP: requirements analysis → background materials → style analysis → writing methodology → ...).

"Skills are truly useful. Prompts, workflows, and tool calls are all packaged inside." (2025-10 post 1914)

Lao Huang switched to Claude Desktop + Skills to replace Claude Code in building an AI OS (post 1914, 2025-10). Core insights:

  • File system → Obsidian
  • Connecting Obsidian MCP with Claude
  • Various customization needs → all implemented with Claude Skills
  • Drawbacks: must be online; cannot switch models

Key products (Build in Public):

  • Prompt House (manages prompts through the MCP protocol; Mac version + Web version; sold for over $100 at $9.98)
  • Obsidian MCP (open source, MIT, included in the official MCP collection; post 1522, 2025-07-18)
  • LearnFromYouTube (reflection on the 60-minute limit; post 2032, 2025-11, preparing to go offline)

Substack replaces Obsidian Publish as the main site (post 1699, 2025-09)

"newtype.pro no longer uses Obsidian Publish and has switched to Hugo + Substack. SEO and GEO are too difficult to handle."

Afterward, he bound the domain newtype.pro ($50 one-time / post 1853, 2025-10), and the content was indexed by Perplexity / ChatGPT / qwen (post 2003, 2025-11).

Generation 5: newtype OS Multi-Agent Orchestration (2026-Q1)

Trigger event: Kimi K2 Thinking arrived (post 1965, 2025-11 / classic)

"I started building a personal AI OS based on Kimi K2 Thinking + Claude Code. I had talked about it for a long time before, and finally the right moment (model) arrived."

The engineering work is divided into two parts:

  1. Improving external information processing efficiency (Web Clipper clipping → weekly AI-generated reports)
  2. Discussion of key topics (information gathering + logical discussion, invoking Prompt House through Skills)

Architectural core: OpenCode + oh-my-opencode + newtype Profile plugin (post 2257, 2026-01)

Later, he found that the standalone plugin version had limitations (shared configuration files, prone to conflicts), so he built the integrated newtype CLI / newtype OS(post 2425, 2026-02):

npm install -g @newtype-os/cli
newtype

Multi-agent orchestration (post 2276, 2026-01 / classic)

  • Chief(Main Agent, Opus 4.5/4.6): coordinates, thinks, and converses with the user, but does not do the work itself
  • Deputy(Deputy, Sonnet 4.5): receives commands from Chief, assigns them to sub-agents, and filters, summarizes, and returns the results
  • Researcher / Writer / Editor / Archivist / Librarian / Fact-checker: specialized agents
  • Done after sub-agents complete self-evaluation(Confidence score) and return it to Chief; the main agent decides whether to redo the work

Dual modes:

  1. Discussion mode (default): rebut and find faults, prompting the user to provide clearer logic
  2. Execution mode: assign tasks

Memory system (post 2344, 2026-01 / inspired by Clawdbot):

  • Short-term: after the conversation stops, automatically saves .opencode/memory/YYYY-MM-DD.md
  • Long-term: after 7 days, automatically archives to MEMORY.md, and Chief reads it through progressive disclosure (table of contents → summary → full text)

Skills system (built into newtype Profile):

  • Super Analyst (12 frameworks)
  • Super Writer (7-step SOP)
  • Super Editor (precision editing)
  • Super Interview (requirements discovery)
  • Super Workflow (manages SOPs, post 2593, 2026-03)
  • Super Workbench (manages Skills, 2026-05)

Mobile integration: WeClaw + Tailscale (post 2565, 2026-03 / post 2402, 2026-02)

  • WeClaw connects to WeChat (idoubi project),nt wechat setup scan QR code to bind
  • Tailscale self-hosted DERP server, direct connection between iPhone 5G and the Mac at home

Cloud work team — Kimi Claw (posts 2389/2432/2451/2458, 2026-02)

  • Kimi Claw runs in the cloud, regularly collecting AI news, YouTube subtitles, and conducting in-depth research
  • Tool stack: Supadata (subtitle extraction) + GitHub CLI (repository synchronization)
  • Bridging to local: email → Obsidian email plugin; or GitHub → Obsidian Git plugin
  • 8-hour / 24-hour scheduling, runs automatically

Architecture diagram:

graph LR
  USER[User] --> CHIEF[Chief<br/>Main Agent / Opus]
  CHIEF --> DEPUTY[Deputy<br/>Assistant / Sonnet]
  DEPUTY --> RES[Researcher<br/>Gemini large context]
  DEPUTY --> WRITER[Writer]
  DEPUTY --> EDITOR[Editor]
  DEPUTY --> ARCH[Archivist<br/>Memory]
  DEPUTY --> LIBRARIAN[Librarian<br/>Researcher]
  DEPUTY --> FC[Fact-checker]

  CHIEF -.mount.-> SKILL[Skills<br/>Super-Analyst/Writer/...]
  CHIEF -.invoke.-> MCP[Prompt House MCP<br/>Obsidian MCP<br/>Tavily/Firecrawl<br/>Sequential-Thinking]
  CHIEF -.read.-> MEM[Memory System<br/>Progressive disclosure<br/>Table of contents→Summary→Full text]

  USER -.WeChat.-> WECLAW[WeClaw]
  WECLAW --> CHIEF
  USER -.iPhone.-> TS[Tailscale]
  TS --> CHIEF

  CLOUD[Kimi Claw Cloud] -.email/Git.-> OBS[Obsidian Vault]
  OBS --> CHIEF

  classDef core fill:#fef3c7,stroke:#92400e,stroke-width:2px
  classDef worker fill:#dbeafe,stroke:#1e40af
  classDef tool fill:#d1fae5,stroke:#065f46
  classDef bridge fill:#fce7f3,stroke:#9d174d
  class CHIEF,DEPUTY core
  class RES,WRITER,EDITOR,ARCH,LIBRARIAN,FC worker
  class SKILL,MCP,MEM tool
  class WECLAW,TS,CLOUD,OBS bridge

3. Three-Layer Architecture (2026 End State)

Lao Huang presented it in post 2462, 2026-02 (featured)the clearest current summary of the architecture:

3.1 Knowledge Base Layer

External materials (historical data, reference materials) + personal materials (past notes).

Current optimal solution:

  • Personal:Obsidian(local-first / md files / iCloud or Dropbox multi-device sync)
  • External:NotebookLM(multimodal / podcasts / Gemini invocation / full Google Drive integration)

Obsidian advanced usage:

  • Obsidian CLI (2026-03 / Obsidian’s most important update in the AI era in 2026) → enables AI programming tools to shift from "blindly traversing the file system" to "operating a knowledge base with indexes and relationship graphs"
  • Add frontmatter properties (type / tag) + wikilinks to each note (Lao Huang manually added them in post 2499, 2026-03, not letting AI do it in bulk, because this is an opportunity for review)

NotebookLM advanced usage:

  • Strict mode (answers only from documents)
  • Multimodal (photo upload / video subtitles)
  • API opened (2025-09 post 1740)
  • Upgrade to Gemini 3 (2025-12 post 2143 / the truly strongest knowledge base app)

3.2 Skills Layer

Skills structurally package the experience in your head.

  • Self-created Skills: Super Analyst / Super Writer / Super Editor / Super Interview / Super Workflow / Super Workbench
  • Repository:github.com/newtype-01/newtype-skills
  • Installation:.skill The file is actually a zip; change the extension and unzip it to use in Claude Code
  • Codex also supports Skills (2025-12 post 2084)

3.3 Real-time Information Layer

OpenClaw and Kimi Claw solve problems that previously had no solution:

  • Scheduled tasks: collect news and YouTube video subtitles every N hours
  • Automation: Bot + Discord/IM push
  • Bridging to local: Email / Git → Obsidian
  • Perplexity Computer $200 tier (2026-03 post 2518) used for investment decisions

4. Complete Tool Stack Chronology (selected, by timeline)

2024

  • 02 Getting Started Period: Ollama / LM Studio / AnythingLLM / Reor (local RAG notes) / Obsidian + Copilot
  • 03 LobeChat / QAnything / Open WebUI / Docker Tried all of them
  • 04 Phidata(RAG + Agent)/ MaxKB Enterprise edition
  • 05 LlamaIndex / RAGFlow Trial
  • 07 GraphRAG(the pain of $11/question + successful heavily modified qwen2 + nomic) / PDF→MD(paid Mathpix → Marker → MinerU → markitdown)
  • 08 Cursor went viral ("build a no-code Chrome summarization plugin in 10 minutes") / Fine-tuning introduction(Unsloth + Colab + Llama 3.1)
  • 09 newtype.pro Blog(hand-built with v0 + Cursor) / ComfyUI + Flux + LoRA
  • 10 Apple AI / Claude controlling computers / NotebookLM Chinese / Learn about / Brave built-in Ollama
  • 11 Web Clipper replacing Omnivore / Mac mini M4 cluster(4 + 1 MBP M4 Max running Qwen 2.5 Coder 32B = 18 tok/s) / Thinking-Claude("performative thinking") / GitHub Copilot supports Claude
  • 12 Self-built iOS Chatbot(Cursor + Tongyi Qianwen Max API, because ChatGPT was unstable in mainland China) / Mac mini M4 for 7K on Xianyu(Hong Kong/Macau channel) / markitdown(Microsoft)

2025

  • 01 Jetson Nano Super $249 / GB10 / DGX Spark launched in May / DeepSeek-R1 1.5B in PocketPal(iPhone 12 mini)
  • 02 Unsloth GRPO fine-tuning(qwen2.5 3B / Colab T4 for 1 hour / 9.9 vs 9.11 testing) / Cherry Studio / ChatWise / Jina DeepResearch / ima knowledge base(Tencent / WeChat File Transfer Assistant) / Grok 3 / Full-capability DeepSeek deployment cost / Deep Research + Cursor + Obsidian trio(post 854) / MBP M4 Pro 48G(nano screen) / Claude 3.7(No. 1 in AI Coding)
  • 03 Manus goes viral / DeepSeek-V3-0324 dirt-cheap / Cline self-built Obsidian MCP / Replicate Flux MCP / PPT MCP / Blender MCP / Gemini 2.5 Pro god-tier(scanning code repositories + article suggestions) / Basic Memory(MCP creating notes to local md) / Second Me(RAG + GraphRAG + fine-tuning → digital avatar) / Milvus + MCP local vector database
  • 04 Grok Workspace(knowledge base + Google Drive) / Cursor + .cursorrules used as knowledge base / Dify MCP(bidirectional MCP) / BiliNote(AI video notes) / After MCP comes AEO(Agent-oriented optimization) / GPT-image-1 API / US company registration Stripe Atlas $500
  • 05 NotebookLM Chinese podcast(post 1180) / Perplexity Comet early version / DeepWiki(deep dive into GitHub projects) / Obsidian Publish + domain name($50, later switched to Hugo) / Prompt House v1 development + launch (prompthouse.vercel.app) / Claude 4 Sonnet(actively using Sequential Thinking)
  • 06 Mac version of Prompt House(Electron + Stripe + Resend + Supabase / $9.98)/ Claude Code for VSCode / GitHub Copilot gray-market relay / Claude Code $100 tier(paid for itself in two days)
  • 07 Going overseas on Reddit(Obsidian MCP viewed 5K+) / Anthropic DXT file(one-click MCP installation) / Stripe / Vercel documentation is LLM-friendly(GEO)/ Canceled Cursor and switched to Claude Code / n8n + MCP Trigger(workflow turned into MCP) / Spokenly speech-to-text / Claude Code SDK is the most overlooked killer feature
  • 08 Claude Code Sub-agents model tiering / CLAUDE.md > 40K split / Claude Code hooks(sound effects) / Sora2 invitation code / WeChat public account paid articles + going overseas on Reddit / GitHub Pro+ Claude integration / GEO(llms.txt + script tags) / Substack replacing Hugo
  • 09 US Capital One credit card / Gemini Deep Research enhanced / Codex replacing Claude Code / Life OS launched / NotebookLM API / Obsidian Velocity theme / Google Fi registration / eSIM(Hong Kong version iPhone 17 Pro) / Qwen3-8B in iPhone(12 tok/s)
  • 10 Claude Skills appeared(2025-10-19 post 1880) / Super Analyst Skill first one / Codex fully replaces Claude Code / ChatGPT Atlas / Perplexity Comet / Sub-agents directory site subagents.cc / First paid WeChat public account article, 24,500 followers
  • 11 Kimi K2 Thinking → AI OS launched(post 1965) / Substack 200 subscribers + 10 paid / Substack content indexed by AI / OpenRouter glm-4.6 running Skills tests → Kimi K2 is best / Codex supports Skills
  • 12 Codex Pro $200 / OpenCode black cards $200 sold out / NotebookLM + Gemini 3 = strongest knowledge base / Gemini adds NotebookLM notebooks / Johnny.Decimal organization method / GEM (Market Insights / Gemini internalization workflow) / US tax and accounting (Taxually + Stripe)

2026

  • 01 OpenCode + oh-my-opencode multi-Agent / newtype Profile plugin v1 / OpenCode black card 200 / Codex used within OpenCode / newtype CLI integrated version released (2026-02-16) / Memory System / progressive disclosure / Knowlery(Obsidian plugin made by Jay, Knowledge Cookery) / Personal Intelligence Gemini(reading Gmail/Maps/Photos) / AI version of Chrome(Gemini in Chrome)
  • 02 Kimi Claw cloud / WeClaw WeChat / self-built Tailscale DERP / OpenClaw + Discord bot / Blink iOS terminal / Spokenly + Typeless voice input (90 yuan per month, 3.2% recognition error rate) / Chrome split view
  • 03 Obsidian CLI(letting AI operate Obsidian through IPC / "blind traversal → index retrieval + knowledge graph") / Perplexity Personal Computer / Perplexity Computer $200 / Obsidian proactively adding frontmatter and wikilinks(2499)/ wiki.newtype.pro established(Karpathy LLM Wiki concept) / WeClaw connected to newtype OS
  • 04 Opus 4.7 GitHub Copilot promotion / GitHub suspends Copilot sign-ups / OpenCode GO plan domestic models / Codex takes over SEO/GEO automation / wiki.newtype.pro goes live / OpenRouter + U.S. credit card = a new route to Anthropic models
  • 05 Knowledge Planet Skill / HyperFrame + Remotion animation / Codex plugin marketplace / NotebookLM extracts Skill / Bing backend GEO takes effect(The comparison page Claude Code vs OpenCode was taken by AI for grounding)

5. Complete explanation of newtype OS

5.1 Installation

npm install -g @newtype-os/cli
newtype

The integrated version includes OpenCode + the newtype Profile plugin + WeClaw; the standalone configuration file is located at .config/newtype/newtype-profile.json(does not conflict with the OpenCode side).

5.2 Key commands

Command Function
/connect Connect to model providers (GitHub Copilot / OpenRouter / Anthropic / OpenAI / Gemini)
/agent-models Configure models separately for Chief / Deputy / Researcher / Writer / Archivist / Librarian / Fact-checker
/init Generate AGENTS.md (code of conduct)
/wiki Generate KNOWLEDGE.md / project Wiki (2026-04 post 2667, Karpathy style)
/switch newtype Switch from oh-my-opencode back to the newtype profile
/switch omo Switch back to oh-my-opencode
/memory-consolidate Manually trigger memory archiving (7 days of logs → MEMORY.md)
nt wechat setup Scan the QR code to bind WeChat (WeClaw)
nt wechat start Start the WeChat Bot
nt research "topic" Call Researcher + Fact-checker for research

5.3 Configuration example (.config/newtype/newtype-profile.json)

{
  "agents": {
    "chief": { "model": "github-copilot/claude-opus-4-7" },
    "deputy": { "model": "github-copilot/claude-sonnet-4-5" },
    "researcher": { "model": "github-copilot/gemini-2-5-pro" },
    "writer": { "model": "github-copilot/claude-sonnet-4-5" },
    "editor": { "model": "github-copilot/claude-sonnet-4-5" },
    "archivist": { "model": "github-copilot/claude-haiku-4-5" }
  }
}

Model tiering principle: Chief uses the best (Opus), Deputy uses the second-best (Sonnet), Researcher requiring a large context uses Gemini (1-million window), and cost-effective roles such as Archivist use Haiku.

5.4 SOUL.md and the dual-layer persona

  • Bottom layer (hard-coded): capabilities + inner persona — cannot be changed
  • Upper layer (SOUL.md): outer persona — users can modify freely

Lao Huang preset the following for the main Agent (Chief):

  1. Tone: colloquial, direct, and judgment-driven
  2. Ability to challenge: directly rebut when flaws are found
  3. Values: the "super individual" framework (data sovereignty / Go Fundamental / code + narrative / antifragility)
  4. No lecturing: express positions through judgment rather than preaching

5.5 Built-in Skills (current state as of 2026-05)

  • super-analyst: 12 frameworks (first principles / impossible trinity / SWOT / ...), Chief independently selects 1–3 to combine
  • super-writer: 7-step SOP, with 7 built-in style-analysis prompts + 6 writing-methodology prompts
  • super-editor: curated editing
  • super-interview: uncover needs / explore
  • super-workflow: manage SOPs, set before delegating to sub-agents Acceptance criteria
  • super-workbench: manage Skills selection

6. Quick reference for key products / tools

6.1 Obsidian

Why it is infrastructure:

  • A private and flexible writing vault that thinks (2025-04 post 1075 quoting the official definition)
  • Local-First philosophy (corresponding to Notion’s All-in-One Workspace)
  • One note = one Markdown document = extremely AI-friendly
  • iCloud / Dropbox multi-device sync (avoiding the official Sync service / 2026-02 post 2365)

Must-install plugin list (Lao Huang’s current version):

  • Web Clipper(replaces Omnivore, extracts YouTube subtitles, Interpreter automatically generates AI summaries)
  • Surfing(2025-07 post 1452, opens web pages inside Obsidian, uses Claude as a browser)
  • Git(syncs cloud GitHub repositories, bridges with Kimi Claw)
  • Velocity theme(2025-09 post 1739)

Lao Huang’s evolving Obsidian philosophy:

  • 2024: numerous plugins (Copilot / Text Generator / Readwise / Omnivore / Random Note / PDF++ / YTranscript)
  • 2025-08 (post 1596):Almost no longer writes notes in Obsidian, switched to writing in Ulysses + backing up in Obsidian, and deleted all AI plugins
  • 2026-03 (posts 2546 / 2499):Manually add frontmatter and wikilinks(for AI to use, not letting AI do it in bulk — this is an opportunity for review)

6.2 NotebookLM

Why it is the optimal solution for ordinary people (2025-12 post 2147):

  1. Equipped with the strongest model (Gemini 3 / upgraded in 2025-12)
  2. Focused on transforming unstructured data into structured data
  3. It is a knowledge base and can also be called by Gemini as an external knowledge base (2025-12 post 2136)
  4. Has a mobile app and supports photo uploads (2025-12 post 2124)
  5. Included in the $20 Google One subscription

Strict mode vs lenient mode (2025-12 post 2074 / classic)

Application Mode Behavior
NotebookLM Strict Only search from documents; if absent, say "I can’t find it"
Google Drive Gemini Lenient Documents + world knowledge
Gemini APP General Training data + internet access + input, synthesis + moderate inference/filling-in

Limitations:

  • Free version: 50 Sources / Premium version: 300
  • Regional restrictions (some areas)

6.3 Cursor → Claude Code → Codex evolution

Stage Tool Model Reason for cancellation
2024-08 to 2025-06 Cursor Claude 3.5 → 3.7 → Sonnet 4 Quota was too expensive; using Claude Code directly with Claude was cheaper
2025-06 to 2025-09 Claude Code Opus 4 / Sonnet 4 Account banned 4 times, switched to Codex
2025-09 to 2026-04 Codex CLI(Pro) GPT-5 / GPT-5.2 Stable, but Claude models were still better
2026-04 to present newtype OS (integrating OpenCode) Opus 4.7(GitHub Copilot)+ Gemini Pro Flexible multi-model configuration

Key operations:

  • 2025-08 post 1648: CLAUDE.md > 40K, let Claude Code split it itself
  • 2025-08 post 1658:/context view consumption
  • 2025-08 post 1652:/hooks add sound-effect alerts
  • 2026-01 post 2219: VS Code is not suitable for running OpenCode (rendering conflicts + lag) → switched to Warp / later switched to Zed
  • 2026-04 post 2738: use a U.S. credit card to register a new OpenRouter account and buy Credits, through which Anthropic models can be used

6.4 Prompt House (Lao Huang’s first product)

  • Website:prompthouse.app
  • Features: Prompt management + MCP protocol invocation
  • Invocation mechanism: two Tool Calls (list → detailed content, progressive disclosure)
  • Pricing: Mac version $9.98 (permanent + future exclusive Skills recommendations)
  • Tech stack: Vercel + Supabase + Stripe + Resend + Electron (Mac version)

The core value of Prompt House (2025-06 post 1413)

"A Pro User must be using multiple AI tools at the same time and must have a need for Prompt management and invocation."

Advanced usage: Prompts scheduling Prompts (2025-06 post 1381)

Prepare two prompts in advance:

  1. Deep research: require the use of Sequential Thinking + Tavily for at least 3 rounds of decomposition. After completion,use another prompt, "Summary Report," to complete the output.
  2. Summary report: clear logic, sound structure, and each section has conclusions + support.

The model first obtains "Deep Research" to complete the research → then obtains "Summary Report" to complete the output.

"One prompt can lead to another prompt. Combined, they form a lightweight Workflow based on MCP tools + Prompts."

6.5 Reflections on LearnFromYouTube (60-minute limit)

Key Bug (2025-10 post 1891 / highlight)

"AI summary products have one fatal bug:the same content has completely different points of interest for different people..

When you use summary products in depth, you find that they cannot meet your expectations, and what they output are very bland conclusions. It looks like it saves trouble, but very often it fails to distill the key points.

By comparison, if you personally read through it and distill it yourself, the conclusions you reach may be completely different from those given by AI.

How can summaries of learning-oriented content reach 80 points? Or can they only reach 60 points at most, with limited value?

I was still too presumptuous in building products."

→ Lao Huang, 2025-11 Post 2032, decided to take LearnFromYouTube offline and focus on Prompt House SEO.

7. Karpathy LLM Wiki Approach / Building wiki.newtype.pro

7.1 Starting Point (2026-04 Post 2620)

"The article Karpathy shared yesterday about his knowledge system — all md files flat out, then let the LLM compile / analyze / build a wiki. Search does not use obsidian-cli and relies entirely on the wiki index."

Lao Huang’s approach is similar, but he does not let AI handle it fully automatically:Having AI tag all notes and add backlinks consumes a huge number of tokens, so it is done manually.

7.2 wiki.newtype.pro Goes Live (2026-04 Post 2767)

  • Approach: articles are a river, while the Wiki is a map. Articles are responsible for expression, while the Wiki is responsible for accumulation.
  • Key point:comparison pages(such as Claude Code vs OpenCode) are very suitable for GEO, and AI search will use them for grounding (2026-05 Post 2805)
  • Path: Substack (www.newtype.pro) is responsible for continuous publishing + subscriptions; the Wiki site (wiki.newtype.pro) is responsible for structured inclusion + long-tail search + AI citations

7.3 Personal IP Upgrade (2026-04 Post 2772 / Featured)

"When my viewpoints are no longer just scattered articles, but form a navigable, citable, and extensible knowledge base, my personal brand changes from ‘someone who continuously outputs’ to ‘someone who possesses a system of thought.’"

8. Conversational Learning / Learning in the AI Era

8.1 Definition (2024-11 Post 615)

NotebookLM and Learn about (Google) are exemplars of AI-native applications, illustrating the concept of Conversational Learning.

8.2 Practical Path (2025-04 Post 1065 / Featured)

"Fully embrace AI learning, starting with NotebookLM.

At first it may feel unfamiliar, because in the past you read and learned from front to back. Now, however, you do not have to follow that order. For example,start from the automatically generated mind map, and whenever you see anything that interests you, jump directly into it..

To use an analogy, logic is the ‘road network,’ knowledge points are ‘stations,’ and AI is the ‘navigator.’ Starting from any knowledge point, as long as you have AI navigation, you can travel across the entire road network."

8.3 AI Learning Workflow (2025-05 Post 1180)

New workflow sequence:

  1. First look at the mind map automatically generated by NotebookLM mind map to gain a rough understanding of the material’s overall logic and core conclusions
  2. Spend less than 10 minutes listening to the podcast automatically generated by AI automatically generated podcast all the way through
  3. Finally, conduct Q&A with AI (clarification + supplementation)

8.4 Three-Piece Learning Method (2025-02 Post 854)

See §2, third generation.

8.5 Book Summary Prompt (2025-07 Post 1478, shared by Victor)

For the complete prompt, see newtype · Prompt / Context Engineering in Practice Section 5. Core points:

  • Used with Gemini (large context + convert the entire book to md with Markitdown and feed it in)
  • Step 0 Pre-Analysis Framing (type + structure)
  • Step 1 Per-Section Summary
  • Step 2 Main Arguments + Evidence
  • Step 3 Themes & Motifs
  • Step 4 Author Craft / Tone / Style
  • Step 5 Critical Reflection
  • Step 6 Detailed Mind Map (at least 4 levels)

9. Subscription, Network, and Identity Infrastructure

9.1 Network (Iron Rule / 2025-07 Post 1463)

"The network hurdle must be overcome; otherwise, you definitely cannot use AI.

Claude / Claude Code: requires a clean US residential static IP. NotebookLM: available only in selected regions (US market). If you really cannot use them → Gemini + Grok + ChatGPT."

9.2 Gemini + Google One Recommendation (2025-11 Post 2042)

The optimal solution for ordinary users = Google One subscription for $20:

  • The Gemini family bundle (embedded in Gmail / Docs / Drive)
  • NotebookLM Premium (300 Sources)
  • Image generation / video generation quotas
  • Pro models
  • Within China, China Merchants Bank MasterCard / China Construction Bank Visa can be used to pay, with no threshold

Lao Huang’s AI application stack (2026-03 Post 2545):

  • Gemini(primary)
  • Perplexity(investment, serious intelligence gathering)
  • Grok(social media dimension / exclusive X data / investment assistant / straightforward style)
  • Codex / Claude Code(development)
  • NotebookLM(learning)

9.3 Overseas Identity (built gradually, throughout 2025)

  • US SIM card(physical card bought on JD.com, $3/month / 2025-04 Post 1094)
  • UK SIM card(giffgaff empty card on JD.com / 2025-01 Post 725 / £10 Pay as you go)
  • eSIM(2025-09 / Hongcha plan / Austrian number / 2025-10 Post 1898)
  • US company(Stripe Atlas $500 / Delaware LLC / 2025-04)
  • Mercury Bank(took 2 days / 2025-04 Post 1152)
  • Tax ID(EIN / waited 1 month / 2025-06 Post 1368)
  • Capital One secured credit card(2025-11)
  • Apple Developer $99 per year(requires D-U-N-S number / 2025-06 Post 1403)

9.4 Finance and Tax (2025-12 Post 2106)

  • Sales tax: Stripe partner Taxually(from $90/month, initial UK setup $150)
  • Income tax: hire someone else to file
  • AI deputy: ask Gemini whenever you encounter something you do not understand

10. Relationships Among Memory / Notes / Workflow / Skill

10.1 Lao Huang’s Layering (2026-01 Post 2298)

"Why not let sub-agents decide for themselves? Because only the main agent knows information such as user preferences and time constraints, while sub-agents only see their own tasks. Therefore, I choose to let the main agent carry out ‘centralized control.’"

Layer Tools Function Persistence
Task layer OpenCode TodoList Substeps of the current conversation No
Skill layer Super-Analyst / Writer Operational SOPs / methodologies Yes (md files)
Workflow layer Super-Workflow Skill Large-process orchestration + acceptance criteria Yes (md files)
Memory layer .opencode/memory/<date>.md+ MEMORY.md Cross-conversation preferences, decisions, topics Yes (progressive disclosure)
Knowledge base layer Obsidian Vault Notes / articles / video scripts Yes (permanent)
Real-time layer Kimi Claw pushes / RSS News / video subtitles No (persisted only after entering the knowledge base)

10.2 Correspondence with KB Design

With ~/kb/’s system is fully analogous:

  • KB inbox → newtype OS input (web clipping)
  • KB sources → Notes in the Obsidian Vault
  • KB wiki → KNOWLEDGE.md generated by newtype OS through /wiki
  • KB INDEX.md → Index of newtype OS’s KNOWLEDGE.md
  • KB tag → frontmatter attributes

11. Key Failures / Important Counterexamples

11.1 GraphRAG Costing $11 per Q&A (2024-07 Post 360)

Lesson: global RAG costs too much → switched to a locally modified implementation using qwen2 + nomic ($0). Lightweight RAG such as Obsidian Copilot has weak performance and cannot handle large books + complex retrieval.Final conclusion: do not use RAG; switch to MCP direct connection to Obsidian.

11.2 LearnFromYouTube 60-Minute Limit (2025-10 Post 1891)

Lesson: learning-oriented summaries cannot be made like movie synopses.Different people have different points of interest, and an averaged summary cannot distill the key points.. Decided to take it offline and focus on Prompt House.

11.3 Apple Intelligence Is Terrible (2025-06 Post 1354)

"Apple Intelligence is so terrible..."

11.4 ChatGPT / Claude Accounts Banned

  • Claude account banned 4 times (2025-04 / 2025-11)
  • 2026-04 Post 2676: Claude requires identity verification (passport + selfie + camera) → Lao Huang announced switching to OpenCode

11.5 Official Accounts / Zhihu / Video Accounts Are All Unreliable (2024-12 Post 689)

"This is why I don’t use domestic platforms as my home base."

Afterward, Lao Huang’s overseas home base = YouTube + X + Reddit + Medium + Substack + newtype.pro.

11.6 Cursor Cannot Sustain Anthropic Costs (2025-08 Post 1516)

"The Claude model is now hard currency. Whoever provides more model quota will have a hit product (Kiro). Whoever cannot bear the costs and starts contracting will go downhill (Cursor)."

11.7 OpenClaw’s Own Programming Ability Is Weak (2026-02 Post 2392)

"OpenClaw’s programming ability is really too weak... I’ll just use a Discord bot to call OpenCode."

12. Current Recommendations (2026-05 Final State)

12.1 General Users

  • Network: residential static IP + US node
  • Models:Gemini(Google One $20)+ Grok(Twitter Blue Verified)
  • Knowledge base: Obsidian + Web Clipper + NotebookLM
  • Creation: Substack replaces Official Accounts

12.2 Creators / Independent Developers

  • AI OS:newtype OS(npm i -g @newtype-os/cli)
  • Model providers: GitHub Copilot Pro+ (Claude + Gemini) / OpenRouter
  • Knowledge base: Obsidian + Obsidian CLI + NotebookLM
  • Writing: Ulysses (writing) + Obsidian (storage)
  • Content accumulation: newtype.pro (Substack) + wiki.newtype.pro (knowledge map)
  • Browser:Perplexity Comet(Primary, underlining and asking questions)
  • Mobile: Blink (terminal) + Typeless (voice input) + Tailscale (tunnel to home Mac)

12.3 Investment / Serious Intelligence

  • Perplexity Max(Computer and other advanced features, $200 tier)
  • Codex / Claude Code for fact-checking
  • Custom GEM (Market Insights report template, see 2025-12 Post 2135)
  • Tesla / NVDA / cryptocurrency: Grok 4.2 Expert + Perplexity Computer model committee (Opus + GPT-5 + Gemini 3)

13. Related Wikis

14. Immutable Anchors Checklist

7 points Lao Huang repeatedly emphasized and kept unchanged for 26 months:

  1. Your own things should be local-first, do not trust any company
  2. Markdown is infrastructure in the AI era, convert all content to md
  3. Notes = AI Context, let AI understand you
  4. Manually adding tags and wikilinks is an opportunity for review, do not let AI do it in bulk
  5. Learning should be Learning by Doing, not Learning by Reading
  6. If the network issue is not solved, AI cannot be used
  7. Pro Users necessarily use multiple tools, so Prompt management is a real need(Prompt House’s moat)

The wiki was manually distilled from all 1,085 knowledge-base topic posts in the newtype Knowledge Planet from 2024-02 to 2026-05, with a full scan and no sampling. This wiki partially overlaps with newtype · RAG in Practice and Knowledge Base Evolution Partially overlapping — the latter approaches from the perspective of RAG technology, while this wiki approaches from the perspective of knowledge system architecture.

2026-06 Addendum: Knowledge maintenance begins to be handed over to Agents, while judgment remains with humans

This section is an increment to the June text snapshot; the preceding “2026-05 final state” should not be understood as a permanent configuration. In particular, the old experience in Section 14 about “manually adding tags” changed on June 1.

From manually tagging to having Agents follow existing classifications

The author began having Codex batch-supplement Frontmatter every week (structured metadata at the beginning of notes): tags, dates, categories, and authors can be precisely filtered, sorted, and aggregated. He asked it to refer to past practices and reuse tags for the same topics. Automation takes over repetitive maintenance, while classification consistency is still provided by existing standards, rather than creating a temporary set of synonyms for each note.

On June 4, newtype Lite packaged part of the OS’s capabilities into more general Skills for Codex and Claude Code to call. This continues the idea that “models can be replaced, while your own system can accumulate.” However, it was also reported on June 1 that Skill triggering was unstable; on the 8th, in a Q&A about whether local Skills could hide content, the author believed recipients would still come into contact with the files. Trigger reliability, distribution, and content visibility are different design problems and cannot be considered solved merely by packaging something as a Skill.

The materials area must solve “search and use,” not merely provide one more reading entry point

On June 7, the plan was to adopt “profile cards + checklists + downloadable files”: Substack handles reading, the community handles questions and discussion, and the materials area handles structured retrieval and reuse. For example, company cards bring together business breakdowns and financial snapshots, while reports and Skills are available for download. The June 6 plan to connect community identity verification to the website was a conceptual entry point, not proof that the integration was completed.

On the 11th, courses, exclusive articles, and exclusive videos were placed into fixed columns, with the motivation of reducing the cost for new members to find materials; at the same time, he believed ideas were better suited to articles than to hand-holding videos. Readers should be able to distinguish between “the river of updates to follow” and “the map to come back and consult”; otherwise, adding a website only increases maintenance work.

Beyond tools, rediscovering serial bottlenecks

On June 9–10, an AI-native organization was defined as continuously compressing the steps that must be handled one by one by humans, rather than adding more assistants to old processes. The author used the analogy of the steam engine first pumping water and then using that water to drive old equipment: adding new power does not necessarily change old structures. It is an observational lens for designing organizations and does not mean that eliminating all human judgment is better.

In the same month, the pen-and-paper note experience added the other side: outsourced organization is handed to AI, while core thinking such as self-positioning, values, and what one wants and does not want remains in offline paper notebooks. One small notebook records daily matters, and one large notebook carries inspiration and reflection. Slowing down through handwriting is the author’s way of forcing himself to stop, not a universally confirmed conclusion from a cognitive experiment this time.

Material boundaries of community tool recommendations

The Card Workspace introduced by the member provides browsing cards by directory, summary previews, virtual scrolling, local search, tag filtering, pinning, and batch operations; Knowlery, for its part, states that version 0.4.0 has been updated and previews “knowledge liveness” in version 0.5.0. Both posts are self-recommendations by developers, and no installation verification has been performed. They can serve as candidate leads when card-style browsing or knowledge maintenance is needed, but this cannot be used as an endorsement of stability or data security.

Sources and related materials