Documentation

Context Layers

A side-by-side exploration of how humans and AI systems build expertise through progressive context loading

ai-servicers.com

The Universal Principle

Creating expertise requires segmented context delivered in progressive layers.

A new employee cannot be productive on day one with only general skills; they need company context, department standards, project knowledge, and task-specific details loaded sequentially.

Similarly, an AI system cannot be effective with only base training; it needs user preferences, workspace standards, project documentation, and runtime tools loaded in order.

Context organization transforms generalists into specialists.

Six Progressive Layers

0
Baseline
1
Broad
2
Domain
3
Specific
4
Task
5
Collab

Each layer builds on the previous, progressively narrowing from general knowledge to specialized, task-ready expertise.

🤖 Claude Code Example

Claude
Base Model
~/.claude/
User Config
~/projects/
Workspace
.../cicd/
CLAUDE.md
PLAN.md
Task Files
Chat Room
Multi-Agent

💡 The Universal Pattern

Without layered context, you have generalists.

With layered context, you build specialists.

This pattern applies to humans, AI systems, and any entity that needs to become productive in a specialized domain. Context organization is what transforms capability into expertise.

0

Baseline Knowledge

General education and training before specialization
🧑‍💼
College Graduate
General Education

What They Know

  • Programming languages (Python, Java, C++)
  • Algorithms and data structures
  • Software design patterns
  • Academic projects

Capability

Can write general-purpose code, but can't deploy to your infrastructure or follow your standards.

🤖
AI Model
General Training (Claude, ChatGPT, Gemini)

What It Knows

  • Programming languages (Python, Java, C++)
  • Common algorithms and patterns
  • General software practices
  • Public documentation

Capability

Can write general-purpose code, but doesn't know your infrastructure, tools, or project specifics.

1

Broad Context

Company-wide knowledge and user preferences
🏢
Company Onboarding
First Week

What They Learn

  • Company mission and values
  • Common tools (Slack, GitHub, Jira)
  • Security protocols and badges
  • Organizational structure

Example Training

  • Style guide — "All REST endpoints use kebab-case"
  • IT helpdesk — "File a ticket when your VPN drops"
  • Security badge — "Never share door codes"
📁
User Level Context
Session Start • ~/.claude/

What It Learns About You

  • Your personal preferences and settings
  • 20+ reusable how-to guides called "skills"
  • A project manager agent to help coordinate work

Example Skills

  • naming-validator — checks that new services follow naming rules
  • troubleshooting — step-by-step guides for diagnosing problems
  • secrets-security — keeps passwords stored safely, never in code
2

Domain Context

Department standards and workspace configuration
🏫
Department Training
First Month

What They Learn

  • Tech stack (Python, PostgreSQL, Kubernetes)
  • Architecture patterns (microservices, event-driven)
  • Deployment pipeline (CI/CD, staging → prod)
  • Coding standards and network topology
🗂
Project Level Common Context
Workspace Entry • ~/projects/.claude/

What It Loads

  • Full service catalog — what's running and how it connects
  • 8 specialized skills (deployment, networking, databases)
  • 3 role-based agents (Architect, Developer, Security)
  • Coding standards, network maps, and security rules
3

Specific Context

Project-specific knowledge and role assignment
🎯
Role Assignment
First Quarter

What They Learn

  • Assigned service (architecture, APIs, dependencies)
  • Institutional knowledge from team leads
  • Access credentials and permissions
  • Known issues and escalation contacts

Example

"You own the Team Chat Platform. It runs two services — a message gateway and a notification engine. Teammates communicate through channels, can assign work to each other, and there are 267 tests to keep it stable."

📝
Project Specific Context
Project Entry • CLAUDE.md

What It Loads

  • Project CLAUDE.md (architecture, APIs, dependencies)
  • Auto-memory (lessons learned across sessions)
  • Project permissions and tool access
  • Known issues and troubleshooting guides

Example

"You're working on AI Agent Chat. It runs two containers (gateway + agents), uses Matrix for messaging, handles task delegation between agents, and has 267 tests to keep it stable."

4

Task Context

Specific assignment details and runtime tools
📋
Assignment Details
Sprint Planning

What They Receive

  • Business requirements
  • Technical approach and patterns
  • Ticket context (status, priority, blockers)
  • Deployment plan

Tools Learned On-Demand

  • Jira (ticket tracking, status updates)
  • Postman, IDE, debugger
  • Slack/email for notifications
  • Task-specific APIs or libraries
⚡
Run-Time Tool Loading
Work Execution • On-Demand

What It Loads

  • requirements.md (problem statement, constraints)
  • Spec artifacts (spec.md, plan.md, tasks.md)
  • GitLab issue context (labels, status, history)
  • Pipeline configuration (stages, gates)

Tools Loaded On-Demand

  • GitLab API (issue management, labels)
  • MCP tools (code executor, review board)
  • Matrix notifications
  • Skills invoked as needed
5

Collaboration

Sharing expertise and getting diverse perspectives
👥
Cross-Team Knowledge Sharing
Enterprise-Wide

Sharing With Peers

  • Direct messages and team channels (Slack)
  • Task delegation with ticket tracking (Jira)
  • Shared drives and wiki documentation
  • Escalation to senior engineers or leads

Consulting External Experts

  • Stack Overflow / Google (quick answers)
  • Cross-team peer review (fresh perspective)
  • Domain consultant (deep specialized analysis)
🌐
Multi-Agent Architecture
Matrix • File-Based • Multi-Model

Sharing With Peer Agents

  • Matrix visit-rooms (per-agent inboxes)
  • Delegation protocol with status tracking
  • File sharing via MinIO (aichat-files)
  • File-based escalation between agents

Consulting Other Models (MCP Dispatch)

  • Gemini: Internet search, fast analysis
  • Codex: Alternate perspective
  • Claude: Deep analysis, architecture review
🤖 AI Review Board ⚡ MCP Code Executor 💬 Shared AI Chat

Why Multiple Perspectives Matter

Humans

A security engineer spots auth gaps a developer might miss. A cross-team reviewer sees coupling your team accepts as normal. A domain consultant brings deep expertise you don't have in-house.

AI Systems

Gemini finds current CVEs via web search. Codex suggests different API patterns. Claude catches architectural debt through deep analysis. Multiple models catch errors a single model would miss.

File-based sharing enables asynchronous coordination.
Multi-model consultation catches blind spots through diverse "training backgrounds."

Context Creates Expertise

0
Baseline
1
Broad
2
Domain
3
Specific
4
Task
5
Collab

The same layered approach that turns a college graduate into a productive engineer also turns a general AI model into a domain specialist.

AI Review Board MCP Code Executor Shared AI Chat

Context Layers • ai-servicers.com • 2025