Context Engineering

Unified skill for all context engineering patterns in AI agent systems. This skill consolidates context management, compression, degradation detection, and KV-cache optimization into a single entry point with progressive disclosure.

Quick Decision Matrix

Problem	Solution	Reference
Context window filling up	Compression	compression.md
Agent ignoring mid-context info	Degradation Detection	degradation.md
High API costs	KV-Cache Optimization	kv-cache.md
Session state persistence	Session Management	session-management.md

Core Concepts

Context Window Thresholds

Utilization	Action	Technique
<60%	Monitor	No action needed
60-80%	Light compression	Observation masking
80-95%	Aggressive compression	Summarization + compaction
>95%	Emergency	Force session handoff

The Four-Bucket Framework

1. Write: Save non-critical info outside context (scratchpads, files)
2. Select: Pull only relevant context (high-precision retrieval)
3. Compress: Reduce while preserving information
4. Isolate: Separate contexts across sub-agents

Compression Techniques Summary

Technique	Token Overhead	Reduction	Best For
Observation Masking	0%	90-98%	Tool outputs >200 tokens
Summarization	5-7%	60-90%	Mixed content
Compaction	0%	50-80%	Older messages

Quick Pattern - Observation Masking:

Before: 500 lines of tool output (500 tokens)
After:  "See /results/search_20260101.txt" (12 tokens)

Degradation Patterns Summary

Pattern	Symptom	Mitigation
Lost-in-Middle	Info at 40-60% position ignored	Place critical info at start/end
Context Poisoning	Errors compound through references	Require source citations
Context Distraction	Model ignores training knowledge	Quality over quantity
Context Confusion	Incorrect associations	Rigorous context selection
Context Clash	Contradictory information	Establish information hierarchy

KV-Cache Optimization Summary

The Four Principles:

1. Stable Prefix: Never change system prompts across requests
2. Append-Only: Never modify previous messages
3. Deterministic Serialization: Same data = same tokens (sort JSON keys)
4. Explicit Breakpoints: Mark cache boundaries

Session Management Summary

Directory Structure:

.cattoolkit/
├── context/
│   ├── scratchpad.md    # Current thinking/decisions
│   ├── todos.md         # Persistent task tracking
│   ├── context.log      # Session history
│   └── checkpoints/     # State snapshots

Scratchpad Hygiene Rule:
Only update scratchpad for:

• Critical decisions made
• Errors encountered
• Phase changes
• Progress milestones

Attention Manipulation via TodoWrite (Proactive Tracking)

The recitation technique from Manus/Claude Code pushes objectives into recent attention span to prevent "lost-in-the-middle" issues:

The Pattern:

1. Create todo.md at task start
2. Update continuously - Check off completed items, add new ones
3. Recite objectives - Rewrite todo to push global plan into model's recent attention

Why It Works:

• Constant todo rewriting recites objectives into context end
• Avoids "lost-in-the-middle" issues without architectural changes

Implementation:

# Before task
- [ ] Research codebase structure
- [ ] Identify patterns
- [ ] Plan implementation

# After research
- [x] Research codebase structure
- [ ] Identify patterns ← Still visible in recent attention
- [ ] Plan implementation

Best Practice: Update todos after every major tool call to maintain objective visibility.

System Reminders Integration

System reminders combat context degradation through recurring objective injection:

Locations:

1. User messages - System reminders in prompt
2. Tool results - Runtime injections
3. Code execution - Added via scripts

Usage Pattern:

# Add reminder at critical points
echo "Reminder: Focus on authentication edge cases" >> .claude/reminders.txt

Effective Reminders:

• Objective recitation - Reiterate main goal
• Constraint reinforcement - Re-emphasize critical requirements
• Context anchoring - Reference key context elements

Plan Mode Best Practices

Plan mode uses recurring prompts to remind the agent:

Implementation:

• Creates markdown files (PLAN.md) persisted during compaction
• Stored in .cattoolkit/context/
• Accessible via /plan command
• Multiple plan prompts and tool schemas for lifecycle

When to Use:

• Complex tasks requiring 10+ tool calls
• Multi-phase implementations
• When agent appears confused or drifting
• Long-running workflows

Best Practices:

• Create plan at task start
• Update as understanding evolves
• Reference plan in reminders
• Use as context anchor during compaction

Integration Points

Skill	Integration
memory-systems	Long-term memory complements context
agent-orchestration	Each agent manages own context
planning-with-files	Plans stored outside context

Usage

When invoked, this skill will:

1. Assess current context state
2. Identify appropriate technique
3. Apply optimization
4. Generate metrics report

For detailed implementation, see references/ subdirectory.