Context Engineering

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](examples.md) for worked input/output.

---

Workflow

Step 1: Define context budget

| Segment | Token budget | Priority | |---------|--------------|----------|

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

| Source | Type | Typical size | Freshness | Required | |--------|------|--------------|-----------|----------|

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

| When overflow | Strategy | |---------------|----------| | History too long | Summarize older turns / sliding window | | RAG too large | Re-rank and top-k truncate | | Document too large | Chunk and select best chunks | | Tool output large | Summarize or extract fields |

Step 5: Plan caching opportunities

| Content | Cacheable | TTL | Savings | |---------|-----------|-----|---------|

Static system prompts, frequent retrievals, embeddings.

Step 6: Validate

Run Validation checklist.

---

Decision Rules

| Condition | Action | |-----------|--------| | No context window limit specified | Use conservative default; state assumption | | Required source exceeds budget alone | Recommend summarization, RAG, or larger model | | PII in context sources | Apply redaction before assembly | | Conflicting info across sources | Define precedence rules | | No truncation strategy | Block; overflow will cause failures |

---

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

| Outcome | Recommended Skill | |---------|-------------------| | Design RAG pipeline | rag/rag-architecture-designer | | Optimize prompts | ai/prompt-optimizer | | Chunking strategy | rag/chunking-strategy-advisor | | AI architecture update | ai/ai-solution-architect |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: patonkikh
• Source: patonkikh/APES
• License: MIT

Install and usage instructions live in the source repository linked above.