Minimum Viable Distillation for Cangjie-Skill: 7-Step RIA-TV++ Pipeline Explained

The minimum viable distillation for Cangjie-Skill is the complete RIA-TV++ pipeline applied to a single candidate unit, producing one SKILL.md file with its supporting test-prompts.json and DIGEST.md.

Cangjie-Skill transforms high-value source texts—books, long video transcripts, podcasts—into agent-callable AI skills. This article breaks down the minimum viable distillation pipeline implemented in the kangarooking/cangjie-skill repository, walking through each stage from source text to deployable skill.

What Minimum Viable Distillation Means in Cangjie-Skill

A minimum viable distillation is the smallest working pipeline that still produces a functional skill pack. According to methodology/00-overview.md, every skill must satisfy five invariants: atomicity, traceability, verifiability, evolvability, and user-in-the-loop design. The pipeline achieves this through seven sequential stages.

Stage 0: Adler Analysis

The pipeline begins with Adler analysis, documented in methodology/01-stage0-adler.md. This stage produces BOOK_OVERVIEW.md—a structured document capturing:

• Structure: How the source text is organized
• Interpretation: Core arguments and their meaning
• Critique: Strengths, weaknesses, and limitations
• Application: Where and how the methods apply

The BOOK_OVERVIEW.md serves as the foundation for all downstream extraction. Each skill repository in the wild contains this generated file, ensuring every distilled skill traces back to a systematic source analysis.

Stage 1: Parallel Extraction

From the book overview, five specialized extractors run in parallel to harvest candidate methodological units. These extractors live in the extractors/ folder:

Extractor	Purpose
framework-extractor.md	Identifies systematic frameworks and models
principle-extractor.md	Extracts core principles and rules
case-extractor.md	Captures illustrative examples and case studies
counter-example-extractor.md	Finds boundary conditions and failures
glossary-extractor.md	Defines specialized terminology

This parallel approach ensures comprehensive coverage without premature filtering. The methodology is detailed in methodology/02-stage1-parallel-extract.md.

Stage 1.5: Triple Verification

Not all candidates become skills. The triple verification stage, defined in methodology/03-stage1.5-triple-verify.md, applies three strict filters:

• Cross-domain evidence: Does the method work outside its original context?
• Predictive power: Does it anticipate outcomes better than alternatives?
• Uniqueness: Does it add something not already covered by existing skills?

Only candidates passing all three checks advance to skill construction.

Stage 2: RIA++ Construction

Verified candidates become full skills using the SKILL.md template at templates/SKILL.md.template. This template enforces six RIA++ fields:

- **R** (Reference): Source traceability
- **I** (Interpretation): What the method means
- **A1** (Application v1): Standard use case
- **A2** (Application v2): Edge or creative use case
- **E** (Evolution): How the skill can grow
- **B** (Boundary): When NOT to use this skill

The RIA++ structure, documented in methodology/04-stage2-ria-plus.md, ensures every skill is self-contained yet connected to its origins.

Stage 3: Zettelkasten Linking

Once constructed, skills get integrated into a knowledge graph. methodology/05-stage3-zettelkasten.md describes how INDEX.md records:

• Dependencies between skills (what builds on what)
• Contrasts and conflicts between methods
• The overall "skill map" for navigation

This linking transforms isolated skills into a coherent, navigable system.

Stage 4: Pressure Testing

Before delivery, every skill faces the pressure test. From methodology/06-stage4-pressure-test.md, this stage:

1. Generates test-prompts.json using the schema in templates/DIGEST.md.template
2. Runs trigger-scenario test cases against the skill
3. Enforces a minimum pass rate of 80%

Skills falling below this threshold are rejected or sent back for refinement. The 80% bar ensures reliability without demanding perfection that would block shipping.

Stage 5: Delivery

The final stage, methodology/07-stage5-deliver.md, produces:

• DIGEST.md: Human-readable executive summary
• Installed skill: Copied to the agent's skills/ directory, ready for immediate use

# Logical pipeline flow (pseudocode based on methodology docs)

source_text = load_source("my_book.txt")
book_overview = adler_analyze(source_text)          # → BOOK_OVERVIEW.md

candidates = parallel_extract(book_overview)        # → 5 extractor outputs

verified = triple_verify(candidates)                # filtered to passing units

skill_md = render_template(
    "templates/SKILL.md.template",
    verified[0]
)                                                   # → one SKILL.md

test_prompts = generate_test_prompts(skill_md)      # → test-prompts.json

if run_pressure_test(test_prompts) >= 0.80:         # 80% minimum pass rate

    install_skill(skill_md, "skills/")              # → deployed

    write_digest(skill_md)                          # → DIGEST.md

Key Files in the Minimum Viable Distillation

File	Role
SKILL.md	Core pipeline definition and invariants
templates/SKILL.md.template	RIA++ skill structure
methodology/00-overview.md	RIA-TV++ pipeline overview
methodology/01-stage0-adler.md	Adler analysis stage
methodology/02-stage1-parallel-extract.md	Parallel extraction
methodology/03-stage1.5-triple-verify.md	Triple verification
methodology/04-stage2-ria-plus.md	RIA++ construction
methodology/05-stage3-zettelkasten.md	Skill linking
methodology/06-stage4-pressure-test.md	Pressure testing
methodology/07-stage5-deliver.md	Final delivery
extractors/*-extractor.md	Five parallel extractors

Summary

The minimum viable distillation for Cangjie-Skill consists of seven stages executed once per candidate:

• Adler analysis creates the source overview
• Five parallel extractors harvest candidate methods
• Triple verification filters for quality
• RIA++ construction builds the skill file
• Zettelkasten linking connects to the skill map
• Pressure testing enforces the 80% pass rate
• Delivery produces DIGEST.md and installs to skills/

This pipeline yields a single, atomic, traceable, verified, tested skill ready for agent invocation.

Frequently Asked Questions

What does "minimum viable" specifically mean in Cangjie-Skill?

Minimum viable means running the complete RIA-TV++ pipeline once on a single candidate to produce one working SKILL.md. It is not a subset of stages—it is the full pipeline applied minimally. Skipping any stage (especially triple verification or pressure testing) violates the invariants defined in methodology/00-overview.md and produces an invalid skill pack.

Why is the pressure test pass rate set at 80% rather than 100%?

The 80% minimum pass rate balances reliability against practicality. As implemented in methodology/06-stage4-pressure-test.md, this threshold catches fundamentally broken skills while allowing through methods that may have edge-case limitations. A 100% requirement would block many valuable but imperfect real-world techniques; below 80%, confidence in the skill's utility drops unacceptably.

Can I run just one extractor instead of all five?

No. The parallel extraction design in methodology/02-stage1-parallel-extract.md intentionally uses five specialized extractors because methodological units appear in different forms across source texts. Running only principle-extractor.md would miss case-based methods; using only case-extractor.md would overlook abstract frameworks. The triple verification stage depends on this breadth to make informed filtering decisions.

How does RIA++ differ from standard RIA (Reference, Interpretation, Application)?

RIA++ adds three extensions to the classic model, as defined in templates/SKILL.md.template:

• A2 (second application): Forces consideration of non-obvious or edge uses
• E (evolution): Documents how the skill can improve or adapt
• B (boundary): Explicitly states when the method fails or should not apply

These additions address evolvability and safe deployment—critical for agent-callable skills that may be invoked in unpredictable contexts.

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