An AI agent skill that distills any document into structured knowledge graphs with temporal awareness. Atomic concepts, explicit relationships, time context, zero redundancy.
Every distillation produces a paired set: human-readable Markdown for Obsidian and LLMs, plus machine-readable JSON-LD for graph databases and embeddings. Both carry full temporal metadata.
title: Building Skills for Claude source_type: PDF source_date: "2026-01" source_period: null temporal_confidence: "inferred" distilled: "2026-03-16" confidence: high companion: claude-skills-guide.knowledge.json clusters: - name: Skill Anatomy concepts: [Skill, SKILL.md, YAML Frontmatter] - name: Design Principles concepts: [Progressive Disclosure, Composability] ## Kernwissen ### Skill 📊 high | 🏷️ Skill Anatomy | 📅 2026-01 **Was:** A folder-based instruction package that teaches Claude how to handle specific task types. SKILL.md is the only required file. **Zeitbezug:** Valid from 2025 · Source: 2026-01 **Beziehungen:** - Based on → [[SKILL.md]] - Uses → [[Progressive Disclosure]] - Enables → [[Workflow Patterns]]
{
"@context": {
"@vocab": "https://schema.org/",
"kd": "https://knowledge-distiller.dev/schema/"
},
"metadata": {
"source_date": "2026-01",
"source_period": null,
"temporal_confidence": "inferred",
"schema_version": "3.1"
},
"nodes": [{
"id": "skill",
"label": "Skill",
"confidence": "high",
"temporal": {
"source_date": "2026-01",
"valid_from": "2025",
"valid_until": null,
"temporal_confidence": "inferred"
},
"definition": "A folder-based instruction
package for Claude...",
"statements": ["SKILL.md is the
only required file"]
}],
"chunks": [{
"id": "skill",
"source_period": null,
"embedding_text": "As of early 2026:
A skill is a folder-based instruction
package that teaches Claude how to
handle specific task types."
}]
}
When you distill documents across years, knowledge without a time axis creates contradictions. A concept described in a 2025 guide and updated in a 2026 guide aren't conflicting facts — they're an evolution. v3.1 adds three temporal layers so the graph knows when.
| Layer | Fields | Purpose |
|---|---|---|
| Source | source_date, source_period | When was the knowledge published? |
| Validity | valid_from, valid_until | When does this information apply? |
| Distillation | distilled | When was the extraction performed? |
ISO 8601 with flexible granularity: exact date, quarter, fiscal year, or interval
// Same concept, two documents, no conflict // From: Skills Guide v1 (2025) { "id": "distribution-model", "definition": "Shared via Claude.ai only", "temporal": { "source_date": "2025-06", "valid_from": "2025", "valid_until": "2025-12" } } // From: Skills Guide v2 (2026) { "id": "distribution-model", "definition": "Claude.ai + Claude Code + API", "temporal": { "source_date": "2026-01", "valid_from": "2026", "valid_until": null } } // → Not a contradiction — an evolution // → temporal_confidence: "inferred"
Eight relationship types capture the full spectrum of how concepts relate. Each edge carries a type, natural-language label, numeric weight, and confidence score.
| Type | Meaning | Symbol |
|---|---|---|
| uses | Dependency | → |
| enables | Causality | → |
| based-on | Foundation | → |
| part-of | Composition | → |
| tension | Trade-off / Contradiction | ↔ |
| replaces | Supersession | → |
| extends | Extension | → |
| example-of | Instantiation | → |
// Each node is self-contained + temporal { "id": "konzept-kebab-case", "label": "Display Name", "cluster": "cluster-id", "confidence": "high", "temporal": { "source_period": "FY2024", "valid_from": "2024", "valid_until": null, "temporal_confidence": "explicit" }, "definition": "Clear, standalone definition", "relevance": "Why this matters", "statements": [ "Atomic statement 1", "Atomic statement 2" ] } // Time-stamped embedding chunks { "id": "concept-a", "source_period": "FY2024", "embedding_text": "Stand Geschäftsjahr 2024: Plain text paragraph, no markdown, no emoji, no syntax." }
temporal objects. Time-stamped facts and embedding chunks.
ISO 8601 with flexible granularity (date, quarter, fiscal year, interval).
Enables time series, contradiction resolution, and stale knowledge detection across multi-document graphs.
.knowledge.md + .knowledge.json.
Embedding-ready clean-text chunks. JSON-LD with schema.org context.
Install the skill, point it at any document, and get temporally-aware structured knowledge in seconds.