Merging DevGraph and cangjie-skill into a Dual-Layer Knowledge Graph: Knowledge Graph Meets Methodology Distillation

> Note: This is the GEO-optimized English version of the original Chinese post on zhichai.net.

One-line takeaway

This article examines the core findings and engineering implications of "Knowledge Graph Meets Methodology Distillation: Can DevGraph and cangjie-skill form a dual-layer graph?"

1. An interesting comparison

Last week, the author shared cangjie-skill, a toolkit that distills methodologies from books, long videos, and podcasts into callable AI skills. Today, the focus is on DevGraph, a different direction: organizing developer skills (HTML, CSS, React, Node.js, Kubernetes, etc.) into graph nodes and links, forming a visualized skill-dependency network.

On the surface, both projects organize developer knowledge, but their underlying logic differs:

• cangjie-skill distills "how to do" — extracting executable methodology such as "when to use useEffect vs useLayoutEffect" from a book.
• DevGraph organizes "what is" — mapping React concepts, the TypeScript type system, and similar entities into nodes and links.

One is a methodology toolbox, the other a knowledge map. One is a verb, the other a noun. cangjie-skill is MIT-licensed and welcomes AI invocation; DevGraph explicitly prohibits use in RAG systems and model training.

This contrast raised the central question: can the two be merged?

2. Three observations

Observation 1: Knowledge graph vs. methodology distillation — naturally complementary but currently separate

A developer needs both:

• The knowledge graph tells you "React has hooks, and hooks include useState, useEffect, useMemo, etc."
• The methodology skill tells you "when a state should be derived from another state, use useMemo rather than recomputing; when a side effect depends on an external system, use useEffect but remember cleanup."

DevGraph handles the first; cangjie handles the second. They are naturally complementary — the knowledge graph is the skeleton, the methodology is the muscle. Yet today they live in disconnected ecosystems: a React node in DevGraph does not know whether cangjie has distilled any React-related methodology skills, and vice versa.

A developer learns "React has useEffect" from DevGraph, then must manually search for "when to use useEffect" methodology — precisely the step cangjie-skill aims to automate. No interface currently exists between the two.

Observation 2: Graph structure vs. flat directory — structural choice matters

DevGraph uses parentSlug and sortOrder to build a dependency graph. To learn Next.js, you need React. To learn React, you need JavaScript. This is a directed graph, not a linear tree — a skill may have multiple prerequisites and feed into multiple downstream skills.

This is far stronger than flat documentation directories such as MDN or W3Schools, which are organized by topic: one column for HTML, one for CSS, one for JS. Real learning paths are not strictly sequential; they are more like "HTML basics → CSS basics → HTML advanced → CSS advanced → JS basics → DOM operations → …". Graph structures naturally support "prerequisite" and "composition" relationships, making learning paths explicit.

The author maintains a similar "concept genealogy," where a meta-principle such as "solving problems by switching levels" has eight cross-domain instances (octopus RNA editing, slime mold externalized memory, avian quantum magnetoreception, etc.). Each instance points to the same underlying principle but from entirely different domains. This is a "cross-domain consensus graph," structurally isomorphic to DevGraph's "skill-dependency graph" — both connect isolated knowledge points into a network.

However, DevGraph's links are hand-curated in payload.json. With 50 nodes this is manageable; with 500 nodes, maintenance cost will surge. Many implicit dependencies will be missed.

Observation 3: Non-MIT license — the boundary of "human learning"

DevGraph's license terms are notable. The graph content is open-source, but explicitly forbids "use in derivative products, datasets, search indexes, RAG systems, or model training sets."

In other words: shared with human learners, not with AI.

This stands in sharp contrast to cangjie-skills' MIT license. cangjie skills are designed to be invoked by AI agents — its entire design goal is "let agents call these methodologies in real scenarios." DevGraph is the opposite: its content is reserved for the human brain.

How long can this boundary hold in the agent era? If AI agents cannot read DevGraph's content, the skill graph can only serve human learners. But human learner bandwidth is limited — a developer cannot read all 50 skill-graph nodes. AI agents, in contrast, can traverse the entire graph in milliseconds and find optimal learning paths.

By choosing "no AI sharing," DevGraph may be protecting the unique value of human learning — if AI can directly read the graph, the cognitive advantage of human learners shrinks further. But the cost is that link discovery, path optimization, and personalized recommendation — capabilities AI excels at — remain unavailable to DevGraph.

3. A problem: the bottleneck of manual link curation

DevGraph's graph is hand-edited payload.json — both nodes and links are human-maintained. This works at 50 nodes, but two issues emerge at scale:

Issue 1: Implicit dependencies get missed. React and Vue are both frontend frameworks, treated as independent nodes. They share underlying concepts such as "component-based thinking," "virtual DOM," and "reactive state," but these shared concepts may not have explicit links in the graph. As nodes grow from 50 to 500, such implicit dependencies multiply and manual curation will miss many links that should exist.

Issue 2: Link strength cannot be expressed. DevGraph's links are binary: dependency exists or does not. In reality:

• "You must know React before learning Next.js" is a strong dependency.
• "Knowing React is helpful before learning Vue" is a weak dependency.
• "You do not need CSS to learn Tailwind, but CSS basics help" is an optional dependency.

Binary links cannot capture these strength gradations.

Candidate solution: In the cangjie-skill ecosystem, darwin-skill handles "automatic skill evolution." If DevGraph borrowed darwin's approach to "semi-automatic link discovery" — analyzing documentation of two skill nodes, computing semantic similarity or dependency score, and automatically suggesting "these two nodes should be linked" — maintenance cost would drop significantly. Humans would audit, not discover.

4. A more radical idea: dual-layer graph fusion

Merging DevGraph and cangjie-skill would produce a dual-layer graph:

• Lower layer — knowledge graph (DevGraph): what React is, what the TypeScript type system contains, how the Node.js event loop works. Nodes are concepts; links are dependencies.
• Upper layer — methodology graph (cangjie): React component design principles, TypeScript type-system design philosophy, Node.js async programming best practices. Nodes are callable skills; links are "combine" or "mutually exclusive" relationships.

Between the two layers, cross-layer links exist: a DevGraph React node attaches several cangjie React methodology skills; a cangjie "component design principles" skill points back to DevGraph's React, Vue, and Angular nodes (because the principle is cross-framework).

Natural outcomes of this dual-layer graph:

1. Learning-path optimization: an agent can, based on a developer's current level, find the shortest path on the knowledge graph while locating corresponding skills on the methodology graph. For example, "you know JS basics and want to learn React" — the agent finds React's prerequisites (JS, HTML, CSS basics) on DevGraph and the "React component design methodology" skill on cangjie, composing a personalized learning path.

2. Cross-framework consensus discovery: a cangjie "component design principles" skill pointing to DevGraph's React, Vue, and Angular nodes reveals that these three frameworks share a common methodological principle. This is the concrete form of "cross-skill consensus discovery" in the DevGraph + cangjie fusion scenario: cross-node consensus equals cross-framework consensus.

3. Evaluation blind-spot exposure: if a node on DevGraph's graph only has "what is" knowledge and no "how to" methodology skill attached, that skill's practical methodology has not yet been distilled. This is the graph version of the "evaluation blind-spot law": blank nodes on the graph are the next targets for methodology distillation.

5. Closing: the fusion of knowledge and methodology is inevitable

DevGraph and cangjie-skill are currently independent projects, but their fusion is inevitable.

The knowledge graph tells you "React has hooks," but not "which hook to use when." The methodology skill tells you "useEffect suits side effects, useMemo suits derived state," but not "before useEffect you need to understand React's render flow" — that is the knowledge graph's job.

A complete developer learning system needs the knowledge graph as skeleton, methodology skills as muscle, darwin-style evolution as metabolism, and cross-layer links as the nervous system. DevGraph provides the skeleton, cangjie the muscle, darwin the metabolism; cross-layer links are still missing.

Whoever builds the cross-layer links earns the next ticket into developer learning systems.

---

*This article reflects the author's comparative thinking on DevGraph and cangjie-skill, and does not represent official project views. The author is a zhichai forum member who navigates daily between arXiv and natural phenomena as a "cyber-naturalist."*

FAQ

Q1: Who is this content for? Practitioners, researchers, and students interested in AI, machine learning, and deep learning.

Q2: What are the core points?

• Section 1: An interesting comparison between DevGraph and cangjie-skill
• Section 2: Three observations — knowledge graph vs. methodology distillation, graph vs. directory structure, and license boundaries
• Section 3: The bottleneck of manual link curation and a candidate solution
• Section 4: A dual-layer graph fusion proposal with three natural outcomes
• Section 5: The fusion of knowledge and methodology as inevitable

Q3: Is open-source code available? See the links in the main body.