---

name: context-management version: "1.0.0" category: _core description: "Context Management Skill"

Context Management Skill

Version: 2.0.0 Created: 2026-01-14 Updated: 2026-01-18 Purpose: Manage context files, enforce limits, and continuously improve based on work patterns

Quick Reference

CLAUDE.md Limits:
- Global (~/.claude/CLAUDE.md): 2KB max
- Workspace CLAUDE.md: 4KB max
- Project CLAUDE.md: 8KB max
- CLAUDE.local.md: 2KB max
- Total Active: 16KB (~4K tokens)

Runtime Context:
- %ctx = (current_tokens / 200000) * 100
- Alert: >60% = archive older exchanges
- Critical: >80% = trim to essentials only

---

Part 1: Context File Management

Size Limits (MANDATORY)

File	Max Size	Max Lines	Purpose
~/.claude/CLAUDE.md	2KB	50	Global preferences
Workspace CLAUDE.md	4KB	100	Delegation patterns
Project CLAUDE.md	8KB	200	Project rules
CLAUDE.local.md	2KB	50	User overrides

Validation Command

# Run validation across all repos
./scripts/context/validate_context.sh

# Check single repo
./scripts/context/validate_context.sh digitalmodel

Content Categories

MUST be in CLAUDE.md:

• Mandatory behavioral rules (TDD, batching, etc.)
• Plan mode conventions
• Cross-review requirements
• File organization rules
• Key delegation patterns

MUST be in .claude/docs/ (reference):

• Agent lists and descriptions
• MCP tool reference tables
• Execution workflow diagrams
• Code examples and patterns
• Memory namespace details

Automated Improvement

The skill analyzes past work to suggest improvements:

1. Pattern Detection: Identifies frequently used instructions
2. Redundancy Removal: Flags duplicate content across files
3. Usage Analysis: Tracks which docs are actually loaded
4. Suggestion Generation: Proposes optimizations

---

Part 2: Runtime Context Management

Response Format Rules

Output Constraints

• Tables: Maximum 10 rows, summarize remainder
• Code blocks: Maximum 50 lines, split larger into files
• Lists: Maximum 15 items, aggregate beyond
• Large outputs: Write to .claude/outputs/, return path only

Mandatory Response Ending

STATUS: [complete|in_progress|blocked] | NEXT: [action] | KEY: [metrics]

Prohibited Actions

1. No Echo: Never repeat input data back
2. No Redundancy: Don't repeat previous findings
3. No Over-Explanation: Keep explanations ≤3 sentences
4. No Raw Content: Use file paths instead of pasting
5. No Unbounded Lists: Always cap with "and N more..."

Context Health Indicators

%ctx	Status	Action
0-40%	🟢 Healthy	Normal operation
40-60%	🟡 Elevated	Consider summarizing
60-80%	🟠 High	Archive older exchanges
80-100%	🔴 Critical	Trim to essentials

Recovery Patterns

When context exceeds threshold:

1. Create checkpoint with current state
2. Write to .claude/outputs/session-state.json
3. Clear verbose history
4. Continue with checkpoint reference only

---

Part 3: Continuous Improvement

Learning from Past Work

The skill tracks patterns across sessions:

# .claude/state/context-patterns.yaml
patterns:
  frequently_referenced:
    - ".claude/docs/agents.md"  # 45 loads
    - ".claude/docs/execution-patterns.md"  # 32 loads
  rarely_used:
    - "verbose section X"  # 0 references in 30 days
  suggested_additions:
    - "Add shortcut for OrcaFlex batch processing"
improvement_suggestions:
  - "Move agent list to reference doc (saves 1.5KB)"
  - "Consolidate duplicate file org rules"

Daily Analysis Tasks

1. Size Check: Validate all CLAUDE.md files against limits
2. Pattern Analysis: Analyze git commits for instruction patterns
3. Usage Tracking: Check which docs were loaded
4. Suggestion Generation: Create improvement proposals
5. Report Generation: Output daily health report

Improvement Workflow

1. Validate sizes → Report violations
2. Analyze patterns → Identify redundancy
3. Check usage → Find unused content
4. Generate suggestions → Propose changes
5. Apply approved changes → Update files
6. Commit → Track improvements

---

Part 4: Scripts

validate_context.sh

Location: scripts/context/validate_context.sh

Validates context file sizes and generates report.

analyze_patterns.sh

Location: scripts/context/analyze_patterns.sh

Analyzes git history to identify instruction patterns.

improve_context.sh

Location: scripts/context/improve_context.sh

Applies approved improvements to context files.

daily_context_check.sh

Location: scripts/context/daily_context_check.sh

Runs all checks and generates daily report.

optimize_mcp_context.sh

Location: scripts/optimize-mcp-context.sh

Optimizes MCP configuration across workspace repos by removing bloated servers (flow-nexus, agentic-payments) to save ~8,500 tokens per repo.

# Dry run - see what would change
./scripts/optimize-mcp-context.sh --dry-run

./scripts/optimize-mcp-context.sh --lean

# Apply config with swarm support
./scripts/optimize-mcp-context.sh --with-swarm

Templates used:

• templates/mcp-with-swarm.json - With swarm (adds ruv-swarm)

---

Part 5: Scheduled Execution

Windows Task Scheduler

Task: ContextManagementDaily Schedule: Daily at 6:00 AM Action: scripts/context/daily_context_check.sh Output: .claude/reports/context-health-YYYY-MM-DD.md

Daily Check Includes

The daily_context_check.sh script runs:

1. validate_context.sh - Size validation
2. analyze_patterns.sh - Pattern analysis (7 days)
3. improve_context.sh --dry-run - Improvement suggestions
4. optimize-mcp-context.sh --dry-run - MCP optimization check
5. Repository status table generation

Setup Command

# Run as Administrator
schtasks /create /tn "ContextManagementDaily" /tr "D:\workspace-hub\scripts\context\daily_context_check.sh" /sc daily /st 06:00

---

Version History

• 2.0.0 (2026-01-18): Add file management, continuous improvement, scheduled tasks
• 1.1.0 (2026-01-17): Add recovery patterns, auto-archive triggers
• 1.0.0 (2026-01-14): Initial context management skill