temple-generator（寺庙发电机）

Temple Generator

Generate a 3D interactive knowledge visualization from any Obsidian vault. The output is a single HTML file (Three.js) with concentric entity rings, audio, discovery mechanics, and multi-scale semantic zoom.

When to Use

• User wants to visualize any Obsidian vault as a 3D knowledge map
• User wants to compare two vaults/document sets visually
• User wants to regenerate the temple from scratch with fresh vault analysis

Architecture

Two-part system:

1. Generation pipeline (this skill): discovers structure, names it, scores confidence, exports a scene package
2. Runtime renderer (template): handles navigation, transitions, audio, discovery

Pre-generate meaning. Runtime-render experience.

Workflow

Step 1: Scan the Vault

Run python3 ~/.claude/skills/temple-generator/scripts/extract_entities.py <vault_path>.

This produces vault-scan.json with:

• Files: path, title, tags, outgoing links, backlink counts, word count, folder, frontmatter
• Graph: adjacency list with bidirectional link counts
• Centrality: degree centrality per node
• Clusters: detected groups of tightly linked notes

Step 2: Read the Scan + Sample Notes

1. Read vault-scan.json
2. Read the top ~20 nodes by centrality (first 100 lines each)
3. Read references/classification-guide.md for entity type heuristics
4. Read 3-5 representative notes to calibrate the vault's "voice" (formal/informal, domain jargon, language)

Step 3: Classify Entities

Using references/classification-guide.md, assign each significant node to an entity type. Maintain two vocabularies:

• canonical: neutral labels for portability (anxiety-management, fermentation-process)
• poetic: mythic/art labels for the installation (The Ferment Gate, The Cortisol Throne)

Target counts per type (adjust for vault size):

Step 4: Build Abstraction Levels

Levels are confidence-gated — only include a level if the vault supports it.

Level 0 — Entities (always exists): individual nodes with positions, connections, descriptions.

Level 1 — Domains (requires >= 3 meaningful clusters): groups of related entities. Each domain has:

• canonical + poetic name
• member entity keys
• centroid position (weighted average of member positions)
• representative exemplar (most central member)
• description (1-2 sentences in vault voice)
• confidence score (0-1)

Level 2 — Axes (requires >= 2 interpretable opposing pairs): fundamental tensions. Each axis has:

• two poles with names and descriptions
• member domains per pole
• axis description
• confidence score

Level 3 — Comparison (requires two vaults + sufficient alignment): shared/unique analysis.

Read references/merge-algorithm.md for dual-graph logic.

Step 5: Generate Scene Package

Follow the schema in references/entity-schema.md to produce temple-data.json.

Include:

• entities: all classified nodes
• levels: abstraction layers with zoom thresholds
• mappings: entity → domain → axis crosswalks
• comparison: (if dual-graph) shared/unique/alignment data
• audio: motif hints per type and level
• style: poetic vocabulary, intro text, color palette, layer definitions
• confidence: per-abstraction and per-alignment scores

Step 6: Generate HTML

1. Copy ~/.claude/skills/temple-generator/assets/temple-template.html to the output location
2. If --inline flag: embed the JSON data as const TEMPLE_DATA = {...}; inside the HTML
3. Otherwise: place temple-data.json alongside the HTML

Step 7: Report

Show the user:

• Entity counts by type
• Abstraction levels generated (with confidence scores)
• Top 5 gods/central entities
• Detected tensions
• If dual-graph: overlap percentage and shared domains

Dual-Graph Mode

When --compare vault_path_2 is provided:

1. Scan both vaults independently (Step 1)
2. Classify entities for each vault (Steps 2-3)
3. Run merge algorithm from references/merge-algorithm.md
4. Generate merged scene package with source attribution
5. Template renders shared scaffold with divergence offsets

Quality Guidelines

• Skip trivial notes (daily todos, admin logs, empty stubs)
• Prefer nodes that reveal the vault's actual concerns, not its filing system
• Write in the vault's own voice, calibrated from sample notes
• If a level lacks confidence, omit it rather than fabricating structure
• Each abstraction level must be backed by membership weights, exemplars, and provenance
• "The abstraction hierarchy should be semantic, not just geometric"

Audio Guidance for Template

The template's audio system should respect hierarchical continuity across zoom levels:

• L0 (close): localized, identity-rich — entity whispers and textures
• L1 (medium): regional harmonic beds, cluster pulses
• L2 (far): sparse drones, tension-based tonal movement
• L3 (comparison): stereo/dialogic between two vault voices

Zoom should feel like changing resolution, not changing universes. Motifs relate across scales.