AgentSkillsCN

context-os-basics

构建上下文操作系统的基础模式

SKILL.md
--- frontmatter
name: context-os-basics
description: Foundation patterns for building context operating systems
model: inherit

Context OS Basics

What is a Context OS?

A context operating system is a structured knowledge system where:

  • •AI compounds intelligence over time (never re-teach)
  • •Knowledge persists across sessions
  • •Concepts link to each other (graph, not files)
  • •Two layers separate reusable knowledge from operational docs

The Two-Layer Architecture

Layer 1: Atomic Knowledge Graph (knowledge_base/)

Individual reusable concepts:

  • •Technical knowledge
  • •Business insights
  • •Methodologies and patterns

Each node:

  • •Has structured frontmatter (metadata)
  • •Links to related concepts via [[wiki-links]]
  • •Follows lifecycle: emergent → validated → canonical

Layer 2: Operational Documents (00_foundation/)

Strategic artifacts that COMPOSE Layer 1:

  • •Positioning documents
  • •Messaging frameworks
  • •Process documentation

Key principle: Operational docs REFERENCE atomic concepts, they don't redefine them.

Constitutional Documents

taxonomy.yaml

Defines blessed tags for your system:

  • •Domains (technical, business, methodology)
  • •Node types (concept, pattern, case-study)
  • •Status values (emergent, validated, canonical)
  • •Your custom topic library

Start small. Add tags only after 3+ nodes demonstrate need.

ontology.yaml

Defines how concepts relate:

  • •Relationship types (enables, supports, implements)
  • •Linking requirements (minimum links per node)
  • •Quality thresholds

Evidence-Based Attribution

Every claim needs a source:

  • •[VERIFIED: file:line] - Direct evidence
  • •[INFERRED: logic] - Deduced from evidence
  • •[UNVERIFIABLE] - Cannot confirm (be honest)

Quality standard: If you can't cite a source, don't claim it.

The 4-Phase Framework (Summary)

  1. •Reconnaissance - Understand domain and requirements
  2. •Architecture - Design two-layer structure
  3. •Construction - Build knowledge graph and ops docs
  4. •Validation - Test and document system

Key Anti-Patterns

1. One Continuous Thread

Problem: Single long conversation without writing to files Fix: Write intermediate results to files, preserve attribution chains

2. Context Explosion

Problem: "Need to read all files to answer this" Fix: Use synthesis docs (5% of content answers 95% of questions)

3. Vague Attribution

Problem: "Many customers want this" (no source) Fix: Quantify everything: "14 of 166 (8.4%)" not "many"

The 3-5 Sample Rule

Never automate without sampling first:

  1. •Run broad search
  2. •Sample 3-5 results with context
  3. •Validate patterns are accurate
  4. •Refine and scale

Advanced Patterns (Not Covered Here)

For complex implementations:

  • •Chief of Staff orchestration
  • •Forcing functions and calibration
  • •Team governance patterns
  • •Enterprise taxonomy design
  • •Multi-agent coordination

These require customization for your specific context. Learn more: https://taste.systems