T1K Context Engineering

Context engineering curates the smallest high-signal token set for LLM tasks. The goal: maximize reasoning quality while minimizing token usage across all T1K workflows.

When to Activate

• Designing/debugging T1K agent systems (cook, fix, debug, test, review)
• Context limits constrain subagent performance
• Optimizing cost/latency in multi-agent pipelines
• Building module-scoped context injection
• Implementing agent memory and cross-session persistence

Core Principles

1. Context quality > quantity — High-signal tokens beat exhaustive content
2. Attention is finite — U-shaped curve favors beginning/end positions
3. Progressive disclosure — Load information just-in-time (sessionBaseline → keyword match → references)
4. Isolation prevents degradation — Partition work across subagents
5. Measure before optimizing — Know your baseline

IMPORTANT: Sacrifice grammar for concision. Pass these rules to subagents.

Quick Reference

Topic	Reference
Fundamentals	context-fundamentals.md
Degradation	context-degradation.md
Optimization	context-optimization.md
Compression	context-compression.md
Memory	memory-systems.md
Multi-Agent	multi-agent-patterns.md
Evaluation	evaluation.md
Tool/Skill Design	tool-design.md
Pipelines	project-development.md
Runtime Awareness	runtime-awareness.md
T1K Patterns	t1k-patterns.md

Key Metrics

• Token utilization: warning 70%, optimize 80%
• Multi-agent cost: ~15x single agent baseline
• Compaction target: 50–70% reduction, <5% quality loss
• Cache hit target: 70%+ for stable workloads

Four-Bucket Strategy

1. Write — Save context externally (scratchpads, files, plans/)
2. Select — Pull only relevant context (module scoping, skill activation)
3. Compress — Reduce tokens while preserving info
4. Isolate — Split across subagents (context partitioning)

Runtime Awareness

T1K hooks auto-inject usage awareness via PostToolUse. Thresholds: 70% WARNING, 90% CRITICAL. See runtime-awareness.md for configuration details.

Anti-Patterns

• Exhaustive context over curated context
• Critical info in middle positions
• No compaction triggers before limits
• Single agent for parallelizable tasks
• Injecting all installed module skills into every subagent
• Duplicating hook logic in AI responses

Security

• Never include secrets, tokens, or credentials in context passed to subagents
• Scope injected context to the minimum needed for the task
• Do not log or persist sensitive tool outputs in plans/ or memory files
• Follow rules/security.md for all context-handling operations