Context vs. Prompt Engineering

Prompt Engineering: Writing instructions for single tasks

Context Engineering: Managing all tokens (instructions + tools + data + history) across multi-turn interactions

Key shift: "What to say" → "What information to include"

Core Problem: Context Rot

• More tokens = worse recall/accuracy
• Context is finite with diminishing returns
• LLMs have "attention budget" that depletes

Guiding Principle:

Smallest set of high-signal tokens for desired outcome

Effective Context Components

System Prompts

✅ Simple, direct language at right altitude (not too specific/vague)

✅ Use XML tags or Markdown sections

✅ Start minimal, add based on failures

❌ No hardcoded logic or vague guidance

Tools

✅ Token-efficient, self-contained, minimal overlap

✅ Descriptive parameters, clear purpose

❌ No bloated tool sets or ambiguous choices

Examples

✅ Curate diverse, canonical examples

❌ No edge case laundry lists

Retrieval Strategies

Just-In-Time (JIT):

• Maintain identifiers (paths, links)
• Load data dynamically with tools
• Progressive discovery
• Trade-off: Slower but efficient

Hybrid (Recommended):

• Critical data upfront
• JIT exploration as needed
• Example: Claude Code

Long-Horizon Techniques

1. Compaction

Summarize → restart with compressed version

• Use for: Conversational flow, near limit
• Keep: Key decisions, bugs, recent files
• Discard: Old tool results
• Quick win: Clear deep tool results

2. Note-Taking

Persistent notes outside context

• Use for: Iterative work, milestones
• Examples: To-dos, NOTES.md, progress logs
• Tool: Memory tool (beta)

3. Sub-Agents

Specialized agents → condensed summaries

• Use for: Complex research, parallel work
• Pattern: Main coordinates, subs explore (10k→2k tokens)

Decision Framework

Quick Wins

1. Remove bloated tools/examples
2. Structure prompts with XML/Markdown
3. Use hybrid retrieval
4. Clear old tool results
5. Add memory tool for notes

Core Principle

"Do the simplest thing that works"

Context remains precious - thoughtful curation beats exhaustive inclusion.