Context Engineering

Purpose

Design context assembly strategy: what information enters the model context, in what order, with what budget, and how to handle overflow.

Input: AI architecture, data sources, prompt template, context window limit
Output: Context Engineering specification with budget allocation, assembly pipeline, and truncation rules
Examples: See examples.md for worked input/output.

Workflow

Step 1: Define context budget

Segments: system instructions, retrieved docs, conversation history, tool results, user query.

Total must fit within model context window minus output reserve (typically 20-30%).

Step 2: Inventory context sources

Sources: RAG chunks, user profile, session history, API tool outputs, documents.

Step 3: Design assembly pipeline

Order of assembly (recommended):

1. System instructions (fixed)
2. Critical constraints
3. Retrieved context (ranked)
4. Recent conversation (truncated)
5. User query

Document ranking/scoring for retrieved content.

Step 4: Define truncation strategies

Step 5: Plan caching opportunities

Static system prompts, frequent retrievals, embeddings.

Step 6: Validate

Run Validation checklist.

Decision Rules

Validation

• Token budget allocated per segment
• Budget fits within context window minus output reserve
• All context sources inventoried
• Assembly order documented
• Truncation strategy for each overflow scenario
• Ranking method for retrieved content defined
• Caching opportunities identified
• PII handling addressed if applicable

Anti-patterns

• Stuff everything — no budget, no truncation.
• Wrong order — user query buried under irrelevant context.
• Stale context — no freshness consideration for dynamic data.
• Duplicate content — same info from multiple sources wasting tokens.
• No output reserve — using 100% context for input.

Best Practices

• Reserve 20-30% of context for model output.
• Put most relevant context closest to user query.
• Use semantic ranking for RAG chunks.
• Summarize long histories rather than drop silently.
• Measure context utilization in production metrics.

Output Structure

# Context Engineering: [System Name]

## Model & Budget
- **Context window:** [tokens]
- **Output reserve:** [tokens]
- **Available input:** [tokens]

## Budget Allocation
| Segment | Budget | Priority |
|---------|--------|----------|

## Sources
| Source | Size | Freshness | Required |
|--------|------|-----------|----------|

## Assembly Pipeline
1. [Step]

## Truncation Rules
| Scenario | Strategy |
|----------|----------|

## Caching
| Content | TTL | Expected savings |
|---------|-----|------------------|

## Metrics
| Metric | Target |
|--------|--------|
| Context utilization | <80% |
| Retrieval precision | [target] |

Next Skills