# Production-Grade Prompting with the Anthropic API

## Concepts, Parameters, Streaming, Caching, and a Real Enterprise Example

Prompting in Anthropic (Claude) is not just about writing good instructions.  
In production systems, prompts behave more like **specifications** than conversations.

This article explains **how prompting actually works in Anthropic**, how to design **reliable and compliant prompts**, and how to use **JavaScript** to build real systems—ending with a **production-grade sentiment analysis prompt**.

---

## 1\. How Anthropic Prompting Really Works

Anthropic uses a **message-based API**.  
Each request contains the **entire conversation state**.

Claude does **not** remember anything between API calls.

### Mental model

```javascript
You → send full conversation
Claude → continues from that context
```

Every request must include:

* Who the assistant is
    
* What rules apply
    
* What the user just said
    
* Any images or documents
    

---

## 2\. Messages and Roles (Core Concept)

Anthropic uses three roles:

| Role | Purpose |
| --- | --- |
| `system` | Hard rules, behavior, constraints |
| `user` | Input data or tasks |
| `assistant` | Previous model replies (conversation priming) |

### Example

```javascript
messages: [
  { role: "user", content: "Hello! Only speak Spanish." },
  { role: "assistant", content: "Hola!" },
  { role: "user", content: "How are you?" }
]
```

Claude will answer **in Spanish** because:

* Conversation history establishes the rule
    
* Claude prioritizes consistency
    

### Production rule

> **Put non-negotiable rules in** `system`, not in conversation history.

---

## 3\. System vs User Instructions (Why It Matters)

| Instruction Type | Reliability |
| --- | --- |
| System | Strong, persistent |
| User | Soft, overrideable |
| Assistant | Priming only |

### Production-safe example

```javascript
system: "You must respond only in Spanish."
```

This survives:

* Long conversations
    
* Truncated history
    
* User attempts to override
    

---

## 4\. Temperature: Controlling Randomness

**Temperature controls how predictable Claude’s word choices are.**

It does **not** control intelligence.

| Temperature | Behavior | Use case |
| --- | --- | --- |
| 0.0–0.2 | Deterministic | JSON, analytics |
| 0.3–0.5 | Controlled | Summaries |
| 0.6–0.8 | Creative | Writing |
| 0.9+ | Unstable | Rarely useful |

### Example

```javascript
temperature: 0.0
```

Use this when:

* Output must be parsed
    
* Compliance matters
    
* Decisions depend on correctness
    

---

## 5\. stop\_sequences: When the Model Must Stop

`stop_sequences` tells Claude **when to stop generating**.

```javascript
stop_sequences: ["\nUser:"]
```

Claude stops **immediately** when it reaches that string.

### Common uses

* Prevent role leakage
    
* End JSON cleanly
    
* Stop agent loops
    

### Key rule

> Stop sequences are **guards**, not fixes for bad prompts.

---

## 6\. Streaming Responses (Real-Time Output)

Streaming lets Claude send text **incrementally**.

```javascript
with client.messages.stream(...) as stream:
  for text in stream.text_stream:
    print(text)
```

### Why streaming exists

* Faster perceived response
    
* Better UX
    
* Long outputs
    

### When NOT to stream

* JSON APIs
    
* Tool calls
    
* Strict schema validation
    

### Production pattern

```javascript
let buffer = "";

for (const chunk of stream.text_stream) {
  buffer += chunk;
  process.stdout.write(chunk);
}

// buffer now contains the full response
```

---

## 7\. Multimodal Prompting (Images + Text)

Anthropic allows **image + text together**.

```javascript
{
  role: "user",
  content: [
    { type: "image", source: {...} },
    { type: "text", text: "Count the containers." }
  ]
}
```

### Rules

* `content` must be an array
    
* Image first, instruction second
    
* Images only in `user` role
    

---

## 8\. Prompt Caching and `cache_control`

Anthropic optimizes performance by caching repeated prompt prefixes.

This is good for:

* System prompts
    
* Templates
    
* Policies
    

But dangerous for:

* User data
    
* Transcripts
    
* Images
    
* PII
    

---

## 9\. `cache_control: { type: "ephemeral" }`

```javascript
"cache_control": { "type": "ephemeral" }
```

This means:

> **Use this content once. Do not cache. Do not reuse. Do not persist.**

### When to use it

| Data | Ephemeral |
| --- | --- |
| User input | ✅ |
| Call transcripts | ✅ |
| Images | ✅ |
| Customer feedback | ✅ |

### Why it matters

* Prevents data reuse
    
* Improves compliance posture
    
* Reduces audit risk
    

---

## 10\. JavaScript Example: Safe Prompt with Ephemeral Data

```javascript
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.ANTHROPIC_API_KEY
});

const response = await client.messages.create({
  model: "claude-3-5-sonnet-20240620",
  max_tokens: 200,
  temperature: 0.0,
  system: "You are a deterministic analysis engine.",
  messages: [
    {
      role: "user",
      content: [
        {
          type: "text",
          text: "The service was slow and frustrating.",
          cache_control: { type: "ephemeral" }
        }
      ]
    }
  ]
});

console.log(response.content[0].text);
```

---

## 11\. What Makes a Prompt Production-Grade

A production prompt has:

1. Clear role definition
    
2. Explicit constraints
    
3. Defined output format
    
4. Fail-closed behavior
    
5. Low temperature
    
6. No exposed reasoning
    
7. Safe handling of sensitive data
    

> **Good prompts are specifications, not conversations.**

---

## 12\. Production-Grade Sentiment Analysis Prompt

This prompt follows **all best practices discussed above**.

### Prompt

```javascript
Analyze the sentiment of the following customer message.

Rules:
- Classify sentiment as POSITIVE, NEGATIVE, or NEUTRAL
- Do not include personal data
- Do not infer intent beyond the text
- Limit text fields to 100 characters

Output:
Return ONLY valid JSON.
No explanations.

Schema:
{
  "sentiment": "POSITIVE | NEGATIVE | NEUTRAL",
  "confidence": 0.0-1.0,
  "summary": "string"
}

If sentiment cannot be determined, return:
{
  "sentiment": "NEUTRAL",
  "confidence": 0.0,
  "summary": "Insufficient data"
}
```

---

## 13\. Sentiment Analysis: Full JavaScript Example

```javascript
const response = await client.messages.create({
  model: "claude-3-5-sonnet-20240620",
  max_tokens: 150,
  temperature: 0.0,
  system: "You are a strict sentiment classification engine.",
  messages: [
    {
      role: "user",
      content: [
        {
          type: "text",
          text: "Support was slow and unhelpful.",
          cache_control: { type: "ephemeral" }
        }
      ]
    }
  ]
});

console.log(response.content[0].text);
```

---

## 14\. Final Takeaway

> **Anthropic prompting works best when you treat prompts like contracts: explicit, constrained, deterministic, and safe.**