Context engineering

Skill: Context Engineering

"Context engineering is the art of filling the context window with just the right information, in the right format, at the right time — to reliably achieve the task." — Karpathy, 2025

Decision Tree — What to put in context

What is the agent missing?

├── Task definition unclear → Add: task description + success criteria
├── No examples → Add: 2-5 few-shot examples (input→output pairs)
├── Background knowledge needed → Add: RAG results (top-3 chunks, not 20)
├── Tool results overwhelming → Compress: summarize, keep key values only
├── Long history → Compact: summarize past turns, keep last 3-5 turns raw
├── Multi-modal inputs → Add: image/file after text instruction
├── Agent "forgets" midway → Add: persistent summary at context top
└── Agent loses task focus → Add: explicit reminder at context bottom

6 Context Layers (fill in priority order)

Layer	What	When to include
1. Instructions	System prompt, task definition	Always
2. State	Current task status, memory	When multi-step
3. Background	RAG results, docs	When domain knowledge needed
4. Examples	Few-shot in-context demos	For complex/ambiguous tasks
5. History	Recent conversation turns	Last 3-5 only
6. Tools	Available tool specs	On demand, not all at once

Context Compression Techniques

1. Summarize long history

[CONTEXT SUMMARY — previous N turns]
User asked about X. Agent found Y. Current status: Z.
[END SUMMARY]

[RECENT TURNS — last 3]
User: ...
Assistant: ...

2. RAG result trimming

• Max 3 chunks, not everything retrieved
• Include: source, date, key sentence
• Exclude: full documents, duplicate info

3. Tool result compression

# Instead of dumping full JSON:
Tool result: Found 3 files. Relevant: config.toml (line 42: api_key), .env (line 7: SECRET)

4. Progressive context (for long tasks)

Phase 1 prompt: task + tools (no history)
Phase 2 prompt: phase 1 summary + new task + tools
Phase N prompt: compressed N-1 history + current step

Context Window Budget (8K context model)

System prompt:      500-1000 tokens  (10-12%)
Task description:   200-400 tokens   (5%)
Examples (few-shot): 300-600 tokens  (7%)
RAG/background:     1000-2000 tokens (20%)
Tool specs:         500-1000 tokens  (10%)
Conversation:       remaining        (~45%)
Output buffer:      reserve 1000     (12%)

Anti-Patterns

• ❌ Dump ALL retrieved docs → ✅ Top-3 most relevant chunks
• ❌ Keep full conversation history → ✅ Summarize after 5 turns
• ❌ Put examples at the END → ✅ Put examples just before the task
• ❌ All tools in every prompt → ✅ Inject tools relevant to current step
• ❌ No state management → ✅ Explicit state object in context

For ZeroClaw/Agent Systems

# config.toml — enable context compaction
[agent]
compact_context = true

# Memory: store compressed summaries, not full history
[memory]
backend = "sqlite"
auto_save = true

Trigger hint:smart route for complex context operations (uses GPT-OSS-120B-free for better context reasoning).