deep-dive

What Are the Outputs of Each Stage in the RIA-TV++ Pipeline?

August 13, 2026 kangarooking/cangjie-skill ↗

The RIA-TV++ pipeline generates seven distinct artifacts—from BOOK_OVERVIEW.md in Stage 0 to DIGEST.md and deployable skill directories in Stage 5—that progressively transform raw source material into Claude-Code-compatible skill units.

The RIA-TV++ methodology implemented in the kangarooking/cangjie-skill repository defines a seven-stage Standard Operating Procedure (SOP) that converts textual materials like books and videos into atomic, verifiable skill artifacts. Each stage produces specific deliverables documented in the methodology/ directory that serve as inputs for subsequent processing, ensuring a reproducible workflow from raw content to deployment-ready skills.

Stage 0: BOOK_OVERVIEW.md

According to methodology/01-stage0-adler.md, Stage 0 (整书理解 or Adler Analysis Reading) produces a single contextual anchor:

books/<slug>/BOOK_OVERVIEW.md — Contains the work's type, thesis, structural skeleton, critical analysis, and applicability assessment. This file establishes the global context that every subsequent sub-agent receives when processing the material.

title: 《穷查理宝典》
author: 查理·芒格
thesis: "逆向思维是避免错误的关键"
structure:
  - 基础概念
  - 决策框架
  - 案例分析
critical:
  - 作者时代局限…
applicability:
  - 可转化为决策 skill

Stage 1: Parallel Extraction Candidates

As defined in methodology/02-stage1-parallel-extract.md, Stage 1 runs five parallel sub-agents to extract candidate units. Each sub-agent outputs to a specific file in the candidates/ directory:

candidates/frameworks.md — Decision frameworks and mental models
candidates/principles.md — Guiding principles and heuristics
candidates/cases.md — Concrete case studies and examples
candidates/counter-examples.md — Boundary cases and exceptions
candidates/glossary.md — Terminology and definitions

Each entry contains atomic candidate units with the minimal required fields: id, title, type, source_chapter, source_quote, summary, and tags.

id: f01
title: 逆向思维
type: framework
source_chapter: 第三讲
source_quote: |
  "反过来想,总是反过来想..."
summary: |
  用于在面对决策困境时，先从相反的视角审视问题…
tags: [decision, mental-model]

Stage 1.5: Triple Verification Results

Stage 1.5 (三重验证筛选) filters Stage 1 outputs through three independent validations defined in methodology/03-stage1.5-triple-verify.md:

books/<slug>/verified.md — Candidates that passed V1 (Cross-Domain), V2 (Predictive Power), and V3 (Exclusivity) validations
books/<slug>/rejected/<id>.md — Individual rejection files documenting which validation failed and the specific reasoning

This stage guarantees that only robust, reusable methodology survives to become skills.

id: f01
title: 逆向思维
type: framework
V1_cross_domain:
  passed: true
  evidence:
    - 第 3 讲: 投资决策场景
    - 第 7 讲: 工程设计场景
    - 第 11 讲: 教学方法场景
V2_predictive_power:
  passed: true
  novel_question: "如果面试官问我一个不知道答案的问题该怎么办?"
  derived_answer: "逆问'我最不希望他认为我是什么样的人'..."
V3_exclusivity:
  passed: true
  why_not_common: "常识是'要多想', 逆向思维是'优先反着想'..."

Stage 2: SKILL.md Construction

Per methodology/04-stage2-ria-plus.md, Stage 2 transforms each verified candidate into a structured skill file following the SKILL.md.template:

skills/<skill-name>/SKILL.md — Contains the six-section RIA++ structure (Reading, Interpretation, A1 Past Application, A2 Future Trigger, Execution, Boundary) plus YAML front-matter with activation metadata for Claude-Code

---
name: inversion-thinking
description: |
  何时用：用户在决策纠结、列举正面理由却理不出头绪时；
  何时不要用：纯信息查询场景；
  关键 trigger：...（中英双写）
source_book: 《穷查理宝典》
source_chapter: 第三讲
tags: [decision, mental-model]
related_skills: []
---

# R — Reading

> "反过来想，总是反过来想……"

# I — Interpretation

逆向思维是一种先从相反方向审视问题的思维模型…

# A1 — Past Application

在本书第 3 讲中，作者利用逆向思维决定是否投资…

# A2 — Future Trigger

> 用户在纠结一个决策、列举正面理由却理不出头绪时…

# E — Execution

1. 列出当前决策的所有正向因素。  
2. 逆向询问"最不希望发生的情况"。  
3. 根据答案重新评估正向因素的权重。

# B — Boundary

- 不要在查询 API 参数、天气预报等纯信息请求时触发。
- 作者在第 9 讲警告：逆向思维不适用于情感冲突…

Stage 3: Zettelkasten Links and INDEX.md

According to methodology/05-stage3-zettelkasten.md, Stage 3 converts isolated skills into a navigable knowledge network:

Updated SKILL.md files — Each skill's related_skills field is populated with neighboring skills and relation types (dependency, contrast, composition)
books/<slug>/INDEX.md — A master index listing all skills, grouped by theme, with a Mermaid dependency graph and recommended learning order
books/<slug>/GLOSSARY.md — Consolidated terminology dictionary extracted from candidates/glossary.md


# 《穷查理宝典》 INDEX

## Skills

- **逆向思维** (`inversion-thinking`) – 决策框架
- **能力圈判断** (`circle-of-competence`) – 思维模型

```mermaid
graph LR
  inversion-thinking --> circle-of-competence
  circle-of-competence --> safety-margin

Glossary

See GLOSSARY.md for key terms.


## Stage 4: Pressure Testing Artifacts

Stage 4 validates trigger accuracy and execution reliability using darwin-compatible test suites, as specified in [`methodology/06-stage4-pressure-test.md`](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/06-stage4-pressure-test.md):

- **`<skill-dir>/test-prompts.json`** — A test suite containing `should_trigger`, `should_not_trigger`, and `edge_case` prompts
- **`<skill-dir>/test-results.md`** — Blind-testing report showing pass rates and failure analysis

Failing tests force a return to Stage 2 for skill refinement.

```json
{
  "skill": "inversion-thinking",
  "version": "0.1.0",
  "test_cases": [
    {
      "id": "should-trigger-01",
      "type": "should_trigger",
      "prompt": "我要决定要不要接这个新项目,列了一堆好处但还是没底",
      "expected_behavior": "调用 inversion-thinking, 反问'最不希望发生什么'",
      "notes": "正面场景: 决策纠结"
    }
  ]
}

Stage 5: Delivery Package

The final stage, defined in methodology/07-stage5-deliver.md, produces the consumable outputs:

books/<slug>/DIGEST.md — Long-form narrative digest summarizing the book's essential insights for human readers
skills/<skill-name>/ directories — Fully validated skill directories copied to the user-visible skills/ folder, ready for Claude-Code to load
(Optional) Darwin-skill input — The test-prompts.json files can feed into auto-evolution loops


# 《穷查理宝典》 精华

本书集合了查理·芒格多年投资与人生的思考精髓。核心观点之一是 *逆向思维*——在做决定前先想象最糟糕的结果，以此排除不值得的选择。…

How the Outputs Connect

The pipeline creates a strict dependency chain where each output feeds the next stage:

Stage 0 establishes global context (BOOK_OVERVIEW.md) that all extractors reference
Stage 1 produces raw candidates that Stage 1.5 filters into verified.md
Stage 2 transforms verified units into SKILL.md files with the R-I-A1-A2-E-B structure
Stage 3 links these skills and builds the INDEX.md navigation layer
Stage 4 validates with test-prompts.json, sending failures back to Stage 2
Stage 5 finalizes the DIGEST.md and deployable skills/ artifacts

This flow ensures the RIA-TV++ invariants of atomicity, traceability, verifiability, evolvability, and deliverability.

Summary

Stage 0 outputs BOOK_OVERVIEW.md for establishing global context
Stage 1 outputs five candidates/*.md files with atomic extraction units
Stage 1.5 outputs verified.md and rejected/<id>.md after triple validation
Stage 2 outputs structured SKILL.md files with the six-section RIA++ format
Stage 3 outputs updated skills with related_skills, plus INDEX.md and GLOSSARY.md
Stage 4 outputs test-prompts.json and test-results.md for darwin-compatible validation
Stage 5 outputs DIGEST.md and deployable skill directories in skills/

Frequently Asked Questions

What file format does Stage 2 use for skills?

Stage 2 produces Markdown files with YAML front-matter named SKILL.md. Each file follows the six-section RIA++ structure: Reading (R), Interpretation (I), Past Application (A1), Future Trigger (A2), Execution (E), and Boundary (B). The YAML header contains activation metadata including name, description, and tags that Claude-Code uses to match user queries to the correct skill.

How does Stage 1.5 determine which candidates to reject?

Stage 1.5 applies three independent validations: V1 (Cross-Domain) checks if the concept applies across multiple domains, V2 (Predictive Power) tests if it generates novel answers to unseen questions, and V3 (Exclusivity) verifies it is not merely common sense. Candidates failing any validation move to books/<slug>/rejected/<id>.md with documentation of the specific failure mode and reasoning.

Where does the pipeline store the final deliverable skills?

Stage 5 copies fully validated skills to the skills/<skill-name>/ directory at the repository root. These directories contain the final SKILL.md, test-prompts.json, and test-results.md, making them ready for immediate loading by Claude-Code or optional processing by the Darwin-skill evolution loop.

What is the relationship between INDEX.md and the individual SKILL.md files?

The INDEX.md file generated in Stage 3 serves as the navigation hub for a specific book, listing all derived skills with thematic groupings and a Mermaid dependency graph. It references the individual SKILL.md files, which have been updated in Stage 3 to include related_skills fields linking them to other skills in the network, creating a bidirectional relationship between the global index and local skill files.

Have a question about this repo?

These articles cover the highlights, but your codebase questions are specific. Give your agent direct access to the source. Share this with your agent to get started:

Share the following with your agent to get started:

curl -s "https://instagit.com/install.md"

Add to your MCP client configuration:

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}

Ask your agent:

"Use Instagit MCP to understand how kangarooking/cangjie-skill works."

Works with

Claude Codex Cursor VS Code OpenClaw Any MCP Client

Maintain an open-source project? Get it listed too →