---
name: ck-context-engineering
description: >
  Optimizes prompts, context windows, and AI interaction patterns for better LLM outputs.
  Activate when user says 'improve this prompt', 'optimize context', 'prompt engineering',
  'getting bad AI responses', 'structure prompt better', or 'reduce token usage'.
  Accepts existing prompts, system instructions, and desired output descriptions.
---

## Overview
Applies prompt engineering and context management techniques to improve LLM response quality, reduce hallucinations, and optimize token efficiency across AI-powered workflows.

## When to Use
- Existing prompts produce inconsistent or low-quality outputs
- System prompts need restructuring for a new AI-powered feature
- Reducing context window usage while maintaining output quality
- Designing multi-turn conversation flows
- Creating reusable prompt templates for a team or application

## Don't Use When
- The underlying model needs fine-tuning (prompting won't fix model capability gaps)
- Issue is API/infrastructure, not prompt quality
- Task needs real-time data the model cannot access (use RAG or tool calls instead)

## Steps / Instructions

### 1. Diagnose the Current Prompt

Identify failure modes:
- **Hallucination**: model invents facts → add grounding, cite sources
- **Verbosity**: too much irrelevant output → add format constraints
- **Missed requirements**: model ignores instructions → restructure, use numbered lists
- **Inconsistency**: different results each run → lower temperature, add examples
- **Context overflow**: long conversation loses earlier instructions → summarize / compress

### 2. Apply Prompt Structure Principles

**Clear role definition:**
```
You are a senior TypeScript developer specializing in Next.js.
Your responses are concise, production-ready, and include error handling.
```

**Explicit output format:**
```
Respond ONLY with a JSON object matching this schema:
{
  "summary": string,       // 1-2 sentences
  "steps": string[],       // ordered action items
  "confidence": "high" | "medium" | "low"
}
Do not include explanatory prose outside the JSON.
```

**Constraints before the ask:**
```
Rules:
- Do not use any third-party libraries
- Prefer async/await over callbacks
- Keep functions under 20 lines

Task: Write a function that...
```

### 3. Few-Shot Examples

Add 2–3 examples for tasks with specific output patterns:
```
Input: "user clicked buy"
Output: { "event": "purchase_initiated", "category": "commerce" }

Input: "page loaded"
Output: { "event": "page_view", "category": "navigation" }

Input: "{{USER_INPUT}}"
Output:
```

### 4. Context Window Management

**Reduce noise:**
- Remove redundant instructions
- Trim verbose conversation history
- Summarize long documents before including

**Chunking strategy for long content:**
```
For files > 4000 tokens:
1. Extract only relevant sections
2. Add chunk position metadata: "[Chunk 2/5 of file.ts, lines 120-240]"
3. Request only what the model needs for the current step
```

**System vs user message split:**
- System: persistent role, rules, output format
- User: task-specific content, inputs, context
- Keep system prompt stable; vary user message per request

### 5. Chain-of-Thought for Complex Reasoning

```
Before answering, think step by step:
1. Identify what is being asked
2. List any constraints or edge cases
3. Draft your approach
4. Verify your answer against the constraints
Then provide your final response.
```

For simpler tasks, suppress COT to save tokens:
```
Answer directly without showing your reasoning process.
```

### 6. Temperature and Sampling

| Task Type | Temperature | Notes |
|-----------|-------------|-------|
| Code generation | 0.0–0.2 | Deterministic, fewer errors |
| Classification | 0.0 | Consistent labels |
| Creative writing | 0.7–1.0 | More varied output |
| Summarization | 0.3–0.5 | Balanced accuracy/fluency |

### 7. Evaluate and Iterate

Create an evaluation set:
1. 10–20 representative inputs with expected outputs
2. Run prompt variant against all inputs
3. Score outputs on accuracy, format compliance, conciseness
4. A/B test two prompt versions before deploying

## Notes
- Prompt changes can have non-obvious side effects; always test on a representative sample
- Document prompt versions with rationale (treat prompts like code)
- Use delimiters (`"""`, `---`, `<tags>`) to separate instructions from user content
- Model context limits change — design prompts that degrade gracefully when truncated