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

newtype · Vibe Coding Independent Product Playbook

Huang Yihe's non-technical indie development roadmap from 2024-09 to 2026. Blog → Prompt House → LearnFromYouTube → Super Analyst Skill → newtype Profile/CLI/OS. Tool stack (Next.js + Supabase + Stripe + Cursor + Electron) / Business model (global merchant + tax filing) / Network setup (VPS+VLESS+Reality+SOCKS5) / Go Fundamental design philosophy / AI Coding division of labor

Source collection：Newtype · Published here：2026-09-26 · Note updated：2026-05-16

Vibe CodingProduct BuildingIndependent development

A practical record and experience accumulation of a series of independent products built by Huang Yihe (non-technical background) using the AI tool stack from 2024-09 to 2026-05.

When to open:

  • You want to build independent products but you are not a programmer
  • You are stuck in the endless loop of "which tools to choose"
  • You want to understand the real division of labor in AI Coding (it is not "hand everything over to Cursor")
  • You want to sell products overseas and need to figure out the whole set of company setup / payment / tax filing / network configuration

0. Product Timeline

graph TD
    A["2024-09 newtype.pro Blog<br/>(v0.dev + Cursor)"] --> B["2025-05 YouTube 字幕下载小工具<br/>(Next.js + Supadata + Supabase + Stripe)"]
    B --> C["2025-06 Prompt House v1<br/>Web + MCP"]
    C --> D["2025-06 Prompt House Mac<br/>(Electron + 本地)"]
    D --> E["2025-08 LearnFromYouTube<br/>(多模型分批 + 备用方案)"]
    E --> F["2025-10 Super Analyst Skill<br/>(Claude Skill + MCP)"]
    F --> G["2025-10 Super Analyst Pro<br/>(+ Sequential Thinking MCP)"]
    G --> H["2026-01 newtype Profile 插件<br/>(基于 OpenCode 多 Agent)"]
    H --> I["2026-02 newtype-cli<br/>(npm 包,统一打包)"]
    I --> J["2026-03 newtype OS<br/>(Multi-Agent Content Team CLI)"]

Each step abstracts on top of the previous one. From "building an App" to "building an Agent OS".

1. Unit Economics: The Premise of Non-Technical Indie Development

Core idea (Old Huang answering non-technical development questions, 2025-05):

Idea Content
Weigh your capabilities and resources One person + non-technical + spending a little money → stay away from large/complex projects
Start from a business angle A project does not necessarily need technical moats to succeed. Can you turn technical problems into business problems/arithmetic problems? For example, wholesale APIs and resell them retail
Start from your strengths If marketing, operations, and product are your strengths, use them as the skeleton to architect the entire project, and use technology as execution

Why this is the moat for non-technical indie development:

  • Programmers' strength of "solving problems with the most elegant code" is overrated in indie development scenarios
  • Customers don't pay because the code is pretty; they pay because "the need is hit accurately + the price is reasonable + the experience is smooth"
  • The limitations of non-technical people actually force the discipline of "thinking through the business first, then using AI to write code"

When to use:

  • Come back and read this when you want to build an independent product but are stuck on "Should I learn programming first?" — the answer is think through the business first, Cursor will write the code for you
  • When you hear "the tech stack this project needs is very complex", ask back "Can I solve it by buying off-the-shelf APIs, turning technical risk into money?"

2. Prompt House: Full-Stack Practical Breakdown

Product Positioning:

  1. Prompt management (store prompts you wrote/collected like taking notes, with titles and tags)
  2. Prompt invocation (directly call via MCP from clients such as Cursor / ChatWise / Cherry Studio, two Tool Calls: one to fetch the list, one to fetch the content)

Web Version Tool Stack:

Layer Tool What it does
Framework Next.js Full-stack base
Data Supabase Database + file storage + Auth
Payment Stripe Collect money + refunds + subscriptions
Email Resend Transaction success notifications + marketing emails (free tier 100/day / 3000/month)
OAuth Google OAuth Login
AI writes code Cursor All code generation
Deployment Vercel Edge deployment

Old Huang only needed to do 4 things himself:

  1. Register accounts on each platform
  2. Configure Google OAuth (because it is an external dependency, you need to go to Google Console)
  3. Configure the Supabase database (create tables, set permissions)
  4. Configure Stripe (API Key, Webhook Key)

The rest of the time Cursor was writing code,Old Huang was scrolling short videos.

Why this division of labor works:

  • Cursor cannot help with the 4 things above, because they involve cross-platform configuration + business decisions
  • Writing code is highly pattern-based work, and LLMs can already handle it
  • Free human attention from "typing" and invest it in "business judgment + configuration decisions"

Frontend UI Tool Combination(later upgraded version):

Tool Role
React Project manager and architect (overall logic + data-driven)
Radix UI Feature engineer (unstyled component library, fully controllable, progressive adoption)
Tailwind CSS Visual designer (colors, sizes, layout)
Framer Motion Motion designer (animations + transitions)

When to use: When you want to build a Web product MVP, copy this stack directly; it covers 90% of scenarios.

3. Prompt House Mac: Electron Local Version

Key changes when migrating from the Web version to a local Mac App:

Dimension Web version Mac version
Framework Next.js Electron
Data storage Supabase cloud Local SQLite
User system Email + OAuth Login removed(no need locally)
API Key Old Huang's (backend proxy) User fills in themselves(can't use Old Huang's Key locally, it will blow up)
Model selection Fixed Pull OpenAI Model List + Ollama 11434 port detection
Updates Vercel automatic Connect to Supabase on launch to check + one-click download

License + verification mechanism design(priced at $9.98):

  • Each user has a unique verification code (a License table in the database)
  • The verification code supports 2 devices using it at the same time; a 3rd device automatically kicks off the 1st device (another device table in the database)
  • Payment completed → Webhook triggered → Resend sends email (containing verification code + download link)
  • Download link = dmg file in Supabase Storage (two versions: Intel + Apple Silicon)

Why complexity suddenly rises:

  • The Web version only has "product logic"; the Mac version adds "distribution logic" (download/activation/update), which is the essential difference between SaaS and software sales
  • US compliance requirements (tax filing) further increase complexity, see §8

When to use: When you want to move from Web to Mac/Windows clients, copy this verification model directly (2 concurrent devices + Webhook email + built-in updates), saving a month of tinkering.

4. AI Coding Division of Labor Philosophy: v0 → Claude → Cursor

Earliest practical experience (2024-09 three-tool division of labor):

  1. v0.dev Make UI and Components (pure design draft layer)
  2. Claude Build the framework (architecture layer)
  3. Cursor Fill in details and debug (implementation layer)

Anti-pattern: Hand everything to one tool; the result will most likely be bad.

Why: These AI tools are not yet strong enough to be "end-to-end SOTA"; each has its own strengths. Cursor is strong at writing code but weak at designing UI; v0 produces beautiful UI but is bad at project skeletons.

Advanced (2026-03): GitHub Copilot model provider → hybrid scheme of Claude (planning/logic) + Gemini (large-context Researcher).Different Agents use different models.

When to use:

  • When starting a new project, first figure out "which step uses which tool"; don't just open Cursor right away
  • When you hear "Cursor is no good/Claude is no good/v0 is no good", ask back "At which step is it no good? Would switching tools solve it?"

5. Go Fundamental: The Essence of UI

Mindset: You can make AI-recognizable UI designs without Figma. The essence of all graphics software is just 3 things:

  1. Create objects(shapes / text / images)
  2. Modify properties(color / size / border / shadow)
  3. Manage relationships(layer order / alignment / grouping)

Figma / Sketch / Photoshop / Keynote / PowerPoint all never escape this.

Old Huang's approach: Use Keynote to make UI sketches, then give them to AI to generate code. The reason is he has used Keynote for over a decade, his proficiency is off the charts, and it launches much faster than Figma.

Why this insight is valuable:

  • It flattens the barrier of "I can't use Figma so I can't build products"
  • This insight can be extended to any professional tool: see through the essence → replace it with a tool you are familiar with
  • The ultimate is "tool freedom": your ability lies in design thinking and taste, not in shortcut key proficiency with any specific tool

When to use: Before you want to learn some new software, first ask "What are the 3-5 essential things it does? What tool I already know well can I use to replace it?"

For detailed extension see newtype Mindset · Super Individual Methodology §6 Go Fundamental.

6. LearnFromYouTube: Multi-Model Batching + Fallback Plan

Product: Input a YouTube link → automatically extract the transcript → output an 8-part learning report:

  1. Summary
  2. Key Takeaways
  3. Timestamped Outline
  4. Key Concepts & Entities
  5. Memorable Quotes
  6. AI-Generated Q&A
  7. Actionable Checklist
  8. Formatted Transcript

Engineering Key Points:

Difficulty Solution
The 8-part segmented report is costly Split into 3 parts, generate each with a different model. The first 7 parts use OpenAI/Gemini, the 8th part (Transcript) uses Gemini-2.5-Flash-Lite (cheap and fast)
API Key failure Prepare N Keys for the same provider + OpenRouter as fallback
Long videos Process in batches (no details given, but it is a key point)
High concurrency Queue + async (no details given)

Why this design pattern is reusable:

  • "Split + different models for different tasks" is a universal pattern: use cheap models for low-cost parts, top-tier models for high-quality requirements
  • Multiple Keys + multi-provider fallback is mandatory for any production environment (if a single Key goes down the product is dead)
  • Engineering complexity is concentrated in fault tolerance, not business logic

Two points he is determined to verify (Old Huang's original words):

  1. As long as API costs are low enough + pricing is reasonable, users will be willing to use ready-made tools instead of tinkering themselves
  2. As long as the need is hit accurately, simple AI is also very valuable

When to use:

  • When building AI SaaS products, first break down "which part is expensive / which is cheap", then decide the model combination
  • When designing the API call layer, implement multi-Key + multi-provider fallback from the start; don't wait for an accident to patch it

7. Super Analyst Skill: Reverse-Empowering AI Systems with Skills

What a Skill is (in Claude Desktop): A "capability pack" that AI can actively invoke, composed of prompt SOPs + trigger conditions. After installing it into Claude Desktop, Claude automatically activates it in appropriate scenarios.

Super Analyst v1(2025-10-24):

  • 12 prompts for different scenario analysis frameworks
  • 1 Skill file to trigger selection
  • Prompts are invoked via Prompt House MCP

Super Analyst Pro (upgraded 2025-10-26):

  • Introduced Sequential Thinking MCP into the SOP (making Claude actually do deep thinking, not perform it)
  • Automatically identify problem complexity (Level 1/2/3)
  • Decide the number of search keywords + number of search rounds based on complexity
  • Weight Chinese and English information sources equally
  • Pick at most 4 analysis frameworks, combine them serially or in parallel

Why Skill is newtype's core lever:

  • Skill = encapsulating the experience in your head into a service AI can call
  • Once a Skill is written → AI calls it automatically → no need to manually feed prompts every time
  • Difference from Prompt House (pure prompt storage): Skill is an SOP with trigger logic, not just a template

Old Huang's 2026 advice: "If you can make more than 3 Skills, you not only greatly reverse-empower your own AI system, but also fully sort out your own workflow. This is an extremely cost-effective thing.Start 2026 by making 3 Skills. "

When to use:

  • You find yourself repeatedly feeding Claude/Gemini with the same line of thinking over and over → this is a Skill candidate
  • You just learned an analysis framework (SWOT / First Principles / Impossible Triangle) → encapsulate it into a Skill instead of applying it manually

8. US Company + Tax Filing + Global Merchant

Payment Architecture:

  • The company entity is in the United States (Wyoming)
  • Stripe payment collection → three third-party companies handle automatic tax filing (starting at $90 per month + $150 for initial submission to the UK)
  • Wyoming, USA threshold:US$100,000 or 200 sales transactions before local tax filing is required
  • United Kingdom:Tax filing is required even for a single order(intimidating)

Why use a US entity:

  • Minimal compliance friction for global payment collection
  • Payment providers such as Stripe assume a US structure by default
  • Users naturally have a high level of trust in products with a US address

Incidental costs: company maintenance fee + bank account + Apple Developer ID (for Mac App Store listing/signing) + US static IP (network disguise, see §9)

Lao Huang's summary: "Being a global merchant has upfront costs. It's quite eye-opening and fun!"

When to use: When you are preparing to build products sold to overseas users, skip domestic entities and go directly with US + Stripe + automatic tax filing. Upfront investment ~$1000 + ~$100 per month.

9. Network configuration: VPS + VLESS + Reality + SOCKS5 proxy chain

Problem: Overseas AI tools such as Claude Code will be detected and blocked when accessed from a domestic Chinese network.

Lao Huang's proxy chain solution:

本地 → VPS (VLESS + Reality) → SOCKS5 (美国住宅静态 IP) → 目标网站

What each component does:

Component Function Why
VLESS Lightweight proxy protocol Fast, but has no built-in encryption and is easily detected
Reality Advanced disguise Makes traffic look like normal website visits (e.g., Apple's official website) to bypass DPI inspection
VPS + VLESS + Reality First hop: local encryption to the VPS Hidden using HTTPS traffic characteristics
SOCKS5 (plaintext) Second hop: US residential IP Provides a "residential address" disguise, making target websites think the visitor is a real user rather than a data center

Why not just use a VPN: VPNs use data center IPs, which large model platforms detect and ban. Residential IPs are in ISP pools shared by ordinary users.

When to use:

  • When you want to use Claude Code / Cursor / Anthropic API but are detected and banned, install this setup
  • Do not use free/shared VPNs (their IPs are already banned across the industry)

10. Third passive income stream: Amazon FBA (launched 2025-08)

Approach: Open a store on Amazon + use Amazon Fulfillment by Amazon (FBA) + do not go to the US in person.

Local US product sourcing: Use a platform similar to 1688 ("the US version of 1688").

Why not build a big store: Consistent with the logic of all previous products — one person + leverage the platform + do not challenge giants. Amazon takes a large cut, but it handles warehousing, logistics, traffic, and trust entirely.

Definition of a remote merchant: Work from home in China, run US e-commerce + US software subscription business + invest in US stocks. Physical location has nothing to do with what you do.

When to use:

  • When you want a third income stream that does not depend on time, FBA is a relatively low-barrier entry point (compared to the customer acquisition difficulty of pure To C software products)
  • Beware of the misconception that "you must move to the US"; Amazon Seller Central can be operated remotely

11. Newsletter as the content home base (Substack)

Why Newsletter(3 reasons Lao Huang switched to Substack in 2025-10):

# Reason
1 Resistance against algorithms — Most platforms rely on algorithms to decide what you can see. A Newsletter is guaranteed to be delivered once subscribed, representing certainty
2 Private reading experience — Email is a relatively "private" and "quiet" space. Immersive reading without algorithmic interference
3 "Yours" — Email list is everything. An email list is the most valuable digital asset a creator can take with them. No matter how many followers you have on other platforms, they are not yours

Why it is critical for independent developers / content creators:

  • Not relying on platforms = not affected by platform algorithm changes
  • Direct delivery to user inboxes = conversion rate an order of magnitude higher than social media pushes
  • Email lists can be sold to the next owner (assetization)

Lao Huang's practice:

  • newtype.pro domain + Substack backend
  • Video scripts + community-exclusive videos + courses are all deposited on Substack
  • Content is also posted on domestic Zhishixingqiu (Knowledge Planet) + Video Accounts + Xiaohongshu, but Substack is the main repository

When to use: When you create content with the goal of long-term accumulation rather than short-term viral hits, Substack > Medium > WeChat Official Accounts (because the list is yours).

12. newtype Profile / CLI / OS: from plugin to product

newtype Profile (released 2026-01)

  • Multi-Agent orchestration plugin based on OpenCode
  • GitHub: github.com/newtype-01/newtype-profile
  • Installation method:opencode.json fill in plugin the npm package name newtype-profile

newtype-cli (released 2026-02)

  • Packages OpenCode + newtype Profile plugin to simplify user installation
  • npm install -g newtype-cli Install in one line
  • Configuration file newtype-profile.json in ~/.config/newtype/
  • /agent-models command to enter the configuration menu (no need to edit JSON manually)

newtype-os (renamed 2026-03)

  • npm scope @newtype-os/cli + @newtype-os/plugin
  • Positioning:Multi-Agent Content Team for Agents — not To Human, but To Agent
  • Command example:nt research "MCP vs CLI" --depth deep --focus 协议层 --no-fact-check
  • Accompanying SKILL.md (instruction manual for Agents to read, not for humans)

Why CLI-ification / To Agent-ification:

  • It is easier for Agents to call CLIs than GUIs/APIs (command line = the universal interface in the AI era)
  • AI calls newtype-os → newtype-os calls finer-grained sub-Agents such as Researcher / Editor / Fact-checker → forming a chain of Agents calling Agents
  • Personal IP productization → sold to overseas creators (SaaS model)

When to use:

  • When you want to build "content SaaS for the AI era", this is a concrete, referenceable product form
  • If you want to understand why more and more products are building CLIs instead of SaaS Web, look at Lao Huang's evolution path

13. Lao Huang's practical experience summary (compilation)

Advanced Prompt usage (Post 1381, 2025-06)

Implementing a "prompts scheduling prompts" workflow via Prompt House:

任务开始
  ↓
模型从 Prompt House MCP 拿到 "深度研究" 提示词
  ↓
深度研究提示词要求:用 Sequential Thinking + Tavily 至少 3 轮迭代搜集
  ↓
完成后,模型从 Prompt House 拿 "总结报告" 提示词
  ↓
总结报告提示词要求:逻辑清晰 / 每节有结论和支撑
  ↓
最终输出

Core: One prompt can lead to another prompt. Combined = a lightweight "Workflow based on MCP tools + Prompts", sufficient for complex tasks.

Ultimate personal knowledge base solution (Post 2201, 2026-01)

Your own content → Obsidian (local-first) Other people's content → NotebookLM (performance-first)

Do not trust any cloud for your own notes; put them in Obsidian (one Markdown per note, extremely AI-friendly). Other people's materials (papers / reports / video transcripts) go into NotebookLM; Source-Grounding strict mode prevents hallucinations.

Installing rules for Agents (Post 2462, 2026-02)

Three-layer architecture for Agents integrating into a personal production system:

Layer Content
Knowledge base layer External materials (historical data) + personal materials (past notes)
Skill layer Structurally encapsulate experience in your mind into Skills
Real-time information layer News / new videos / long articles (automated using tools such as OpenClaw)

14. Anti-patterns: don't step into these pits

❌ ✅
"I'll learn frontend completely before building a product" Start building directly; Cursor will write the code, you only need to think clearly
Using one AI tool for the entire process v0 for design / Claude for architecture / Cursor for implementation, each handles its own segment
Running production on a single API Key Multiple Keys + multiple provider fallbacks, from Day 1
Selling overseas using a domestic entity Go directly with a US entity, Stripe + automatic tax filing
Using data center IPs for overseas AI Residential IPs are mandatory, otherwise accounts get banned
Feeding SOPs in your head to AI over and over Write them as a Skill and let AI call it itself
Only posting content on platforms without depositing it in a Newsletter Email list = the most valuable digital asset
Trying to build an "all-in-one product" Single-point breakthrough, penetrate one need point thoroughly

References and extensions

Main references: newtype Knowledge Planet featured posts (2024-09 to 2026-05): 481, 488, 1197, 1207, 1233, 1344, 1358, 1373, 1379, 1381, 1391, 1406, 1412, 1450 (idoubi-ai.pdf), 1492, 1573, 1601, 1615, 1636, 1655, 1669, 1681, 1851, 1875, 1902, 1912, 2076 (NVDA 10 years), 2195, 2201, 2208, 2252, 2425, 2462, 2484, 2509, 2535, 2575, 2619

Companion wiki:

Cross-series related:

  • Wang Kai SOP series (Phase 2 processing) — Multi-Agent browser matrix + Claude Code workflow overview, will be strongly related to §12 newtype-os of this wiki

History

  • 2026-05-16 v1.0 created — wiki #2 extracted from the newtype full post index (Phase 1.1 · 2/12). Source session: Session 2 wiki slices.

2026-06 supplement: Turn content production bottlenecks into tools, then into products

This section comes from June text posts in the newtype community, retaining the time boundaries of the author's experience and plans; accompanying images, attached programs, and videos were not executed or watched in this process; time-saving data is self-reported by the author.

Rough cut automation first outputs a revisable timeline

Huang Yihe separately toolized the most time-consuming A Roll (main speaker footage) rough cut for talking-head videos. The process introduced on June 3 is: extract audio from the original video → local Whisper generates word-level timestamps → align recognized text with the pre-written script → identify re-read segments and long pauses → output an FCPXML editing decision timeline, import into DaVinci Resolve for further fine editing. The original video is not directly modified; the output is an instruction document specifying "which clips to take".

Script alignment solves one more problem than simple silence detection: going back from sentence 5 to sentence 4 may indicate a retake, not just a pause. The example case leaves 0.4 seconds of breathing room before and after clips, aimed at avoiding word truncation and breath cutoff; this was the author's parameter at the time, not a universal standard to copy for all speaking speeds and source materials.

On June 4, the author self-reported a rough cut where a 50-plus-minute original clip generated a 20-plus-minute, 200-plus-segment rough cut, compressing work that previously took one to two hours down to a few minutes. This demonstrates feedback on a personal workflow, not a general performance benchmark. On June 6, newtype-cut.zip was released, with a README for Agents to install and run; on June 8 it was further made into a Skill, where the first call scans the environment and installs missing dependencies after obtaining user consent. Regarding CapCut's support for FCPXML, the author explicitly stated that this came from Perplexity and he had not tested it himself, so it cannot be treated as a compatibility guarantee.

A single methodology can have different delivery layers.

The June 15 product plan for "AI Industry Chain Investment" separated Why (rationale) from How (implementation): the video course was planned to be sold standalone at ¥398, the text column at ¥198, and the print book at ¥60–80; within the community, it was proposed to provide white papers, biweekly portfolio reports, event-driven updates, company dossiers, a data repository, and quarterly retrospective reviews of mistakes. The pricing and release status reflect only the plan at that time. The early June post's notes that "the first episode is done," "the second episode is updated," and the eight-episode plan do not mean the full course has been delivered.

What is worth referencing is the layered delivery: courses explain the framework, continuously updated materials support decision-making, and periodic retrospectives are responsible for correcting errors. Simply reselling video content as articles does not automatically generate the value of the latter two layers. The author also packaged the investment methodology into a Skill, indicating that content and tools can share the same judgment framework; this batch does not include the full text of the Skill, so its investment algorithm cannot be reconstructed.

A creator's consistent choices accumulate more reliably than chasing hot topic schedules.

When reviewing his 178th public video on June 9, the author said that roughly ten of his early videos went viral, and the total count including courses had exceeded two hundred, yet he could not precisely plan the roadmap in advance. What he has stuck to long-term is focusing primarily on Why, touching on How only briefly, and producing little content that chases product launches to introduce What. He believes this allows the audience to build cognition around judgment, rather than only coming to ask for the next operational step. This is a personal content positioning, not a market law that "all tutorials have no value."

Video scripts are synchronized to a WeChat Official Account and Substack, which respectively serve domestic and overseas distribution and provide machine-readable entry points. The author speculates that the content will be read by models, but "being published on the web" does not prove that it has been indexed by any specific model. When choosing distribution carriers, one should distinguish between the act of publishing and the expected outcomes of discovery, subscription, and citation.

2026-06 Addendum: Delivery boundaries of member product cases

The following are self-reports from community members, and have not been installed or retested. Accurately describing what a prototype can do is more valuable than labeling it with a name like "fully automatic."

  • Drawing bubble annotation: A member used Opus in Copilot to produce the first draft and DeepSeek for revisions. Dimensions and tolerances are entered manually, and the output is an Excel inspection sheet, which is then considered for input into an ERP system; this does not automatically parse drawings and generate specifications. Another member's "AI drawing pre-review" is still under exploration, and a short preview is insufficient to prove that it has been delivered.
  • PDF requirement document parsing: The member claims it has been sold to an autonomous driving company, covering four input types based on the presence or absence of an outline and a text layer. After extracting the outline, it allows manual adjustment of slices, exports to Feishu for tree-structured display, and traces back to the original PDF. Existing comments confirm that text can be extracted, there was no search function at the time, and OCR of formulas and tables still suffered from loss issues. The main text is truncated, so the above covers only the visible preview and replies; it cannot be claimed that the complete product has been reviewed.
  • Creamlon Protocol: An Agent service is presented via a GitHub repository, public Issues are used to accept asynchronous orders, the buyer's private repository stores encrypted materials and deliverables, and signed receipts are used to verify the correspondence between source and order. It does not handle payments; WeChat, Stripe, x402, and others are only channels the author proposed that users can connect on their own. Private repositories, encryption, and signatures serve different responsibilities; the statement that "all security is guaranteed by GitHub" is only the publisher's claim and cannot replace an assessment of permission configuration and key management. The project self-describes as being in an early stage, and manual approval and inspection are still required.
  • Leveraging ready-made infrastructure: The author recommends ShipAny to reduce website development time, provides a discount code, and this is a promotional recommendation without a cross-sectional performance evaluation. His own newtype Lite is oriented toward content, and AI Stack toward investment methodology; these are two Skills for different purposes. This batch contains only introductions, not the full code or rule sets.

When to reference: When you already have clear delivery workflows for documents, drawings, or reports and want to productize repetitive steps. Retain manual correction and traceability to original documents before talking about scaling; do not conflate "automatic annotation," "automatic review," and "automatic specification generation" into a single capability.

Sources and related materials