Context Engineering

Context

Even with 1M-2M token context windows, stuffing everything degrades performance ("Lost in the Middle" — Liu et al. 2023). This skill encodes Anthropic's 2025 best practices: dynamic compression, intelligent log truncation, and priority-aware token budgeting.

The core insight: context is RAM, treat it like memory management, not a buffer.

Guidance

Three Tools

from framework.src.context_engineering import (
    ContextCompressor, LogTruncator, TokenEconomy,
    CompressionLevel, BudgetPriority,
)

# 1. Compress long text (LLM outputs, paper abstracts)
compressor = ContextCompressor()
result = compressor.compress(long_paper_abstract,
                              level=CompressionLevel.MEDIUM,
                              focus_query="time series forecasting")
# → AGGRESSIVE for very long, MEDIUM for 4K-20K, LIGHT for <4K

# 2. Truncate training logs (preserve errors + samples)
truncated = LogTruncator(max_lines=200).truncate(training_log)
# → 500 lines → 50 lines, all ERROR/WARN kept, METRIC sampled

# 3. Priority-aware token budget
econ = TokenEconomy()
econ.add_critical(system_prompt, "system")
econ.add_high(current_task, "task")
econ.add(relevant_skill_text, BudgetPriority.MEDIUM, "skills")
econ.add(optional_context, BudgetPriority.LOW, "context")
prompt = econ.assemble()  # auto-drops LOW when over budget

Decision Workflow

Assembling a prompt?
  ↓
1. Are there training logs > 100 lines?
   → LogTruncator first (reduce noise 5-10x)
  ↓
2. Are there paper abstracts / long docs?
   → Compressor + MEDIUM (keep relevant 50%)
  ↓
3. Total still > 8000 chars?
   → TokenEconomy (prioritize, auto-drop LOW)
  ↓
4. Send to LLM

Why This Matters

Real ROI: agy research confirms 20% accuracy improvement on code generation when context is properly engineered.

When to Apply

When to Use

• Any prompt > 4000 chars
• Training logs in context
• Mixing system + task + skills + examples
• "Model keeps ignoring this instruction"
• Multi-turn conversations with growing history

When NOT to Use

• Short prompts (<1000 chars) — overhead not worth it
• When every detail must be preserved (use verbatim)
• Streaming scenarios (compress after stream ends)

Notes

• Combine with Memory Hierarchy (don't compete): Memory decides what enters context; Engineering decides how to fit it
• Measure first: log token counts before/after to validate savings
• Test with edge cases: 50K chars, all-metrics log, mixed priorities
• See also: memory-hierarchy-management, runtime-reflexion

References

• Implementation: framework/src/context_engineering/
• Inspired by: Anthropic Context Engineering Best Practices (2025)
• Research: Liu et al. "Lost in the Middle" (2023), Anthropic prompt caching (2024)