Production Prompts

Write effective prompts for production AI workflows

Production prompts need to be reliable, not clever. This guide covers patterns that work consistently across thousands of runs.

System Messages

The system message defines who the LLM is and how it behaves. Be specific.

Vague (bad):

<system>
You are a helpful assistant.
</system>

Specific (good):

<system>
You are a sales research assistant at a B2B SaaS company. You help sales teams prepare for calls by summarizing prospect information.

Rules:
- Be factual and concise
- Never make up information — if something is unclear, say so
- Focus on business-relevant details: what they do, who they sell to, recent news
- Skip generic company boilerplate
</system>

The specific version produces consistent, useful output. The vague version produces whatever the model feels like.

Constraints Matter

Explicit constraints prevent the most common failure modes:

<system>
You classify leads by buying intent based on their activity.

Constraints:
- Only use the categories: high, medium, low, unknown
- When uncertain between two categories, choose the lower intent
- Never create new categories
- If activity data is insufficient, classify as "unknown"
</system>

Without constraints, the model might invent categories, misclassify edge cases inconsistently, or add explanations you don’t want.

Temperature Settings

Temperature controls randomness. Match it to your task:

Temperature	Use Case	Example
0	Judges, extraction, classification	Evaluator prompts, extracting entities, classifying leads
0.3	Analysis, summarization	Company research summaries, content analysis
0.7	General tasks, balanced creativity	Drafting emails, writing descriptions
1.0+	Brainstorming, creative writing	Generating variations, ideation

For production workflows, err toward lower temperatures. A temperature of 0 means the same input produces the same output — essential for debugging and testing. Judge prompts should always use temperature 0.

[e-mail]

---
provider: anthropic
model: claude-sonnet-4-20250514
temperature: 0
---

<system>
You evaluate company research summaries for a sales team. A good summary must:

1. Mention what the company does (their core product or service)
2. Identify their target market or customer base
3. Include at least one specific, verifiable fact
4. Be 2-4 paragraphs long
5. Not contain any claims that aren't supported by the provided source data
</system>

Structured Output

When you need to parse the result programmatically, use generateText with Output.object() and a Zod schema instead of asking for JSON in the prompt.

Unreliable (parsing JSON from text):

<user>
Extract the company info and return as JSON with fields: name, industry, employee_count.
</user>

Reliable (schema-validated):

const { output } = await generateText({
  prompt: 'extract_company@v1',
  variables: { content: websiteContent },
  output: Output.object({
    schema: z.object({
      name: z.string(),
      industry: z.string(),
      employeeCount: z.number().nullable()
    })
  })
});

The schema is validated server-side by the LLM provider. You get typed output or an error — no parsing surprises.

Keep Schemas Simple

Complex nested schemas confuse models. If your schema has more than 5-6 fields or deep nesting, split it into multiple calls.

Too complex:

const schema = z.object({
  company: z.object({
    name: z.string(),
    details: z.object({
      industry: z.string(),
      subIndustry: z.string(),
      // ... 10 more nested fields
    })
  }),
  competitors: z.array(z.object({
    // ... another complex structure
  }))
});

Better — split into focused calls:

const { output: company } = await generateText({ prompt: 'extract_company@v1', output: Output.object({ schema: companySchema }), ... });
const { output: competitors } = await generateText({ prompt: 'find_competitors@v1', output: Output.array({ element: competitorSchema }), ... });

Few-Shot Examples

Show the LLM what you want with 2-3 examples. This works better than lengthy instructions.

<user>
Classify this lead's buying intent: high, medium, low, or unknown.

Examples:
- "Downloaded pricing page, requested demo, visited 5 pages" → high
- "Read two blog posts, signed up for newsletter" → medium
- "Visited homepage once from a Google ad" → low

Lead activity: {{ activity }}
</user>

Few-shot examples are especially effective for:

• Classification tasks
• Formatting requirements
• Edge case handling
• Tone matching

Common Mistakes

1. No Role Definition

# Bad — no context for the model
<user>
Summarize this company: {{ content }}
</user>

# Good — clear role and purpose
<system>
You create concise company summaries for sales teams.
Your audience is busy sales reps who want key takeaways before a call.
</system>

<user>
Summarize this company: {{ content }}
</user>

2. Unbounded Creativity

# Bad — model will write a novel
<user>
Write about {{ company_name }}.
</user>

# Good — clear bounds
<user>
Write a 2-paragraph company overview for {{ company_name }}.

Paragraph 1: What they do and who they serve.
Paragraph 2: Recent developments or notable achievements.
</user>

3. Asking for Everything at Once

# Bad — too many things, inconsistent results
<user>
Analyze this company: give me an overview, list competitors,
identify their tech stack, summarize recent news, and predict
their growth trajectory.
</user>

# Good — focused request, combine in workflow
<user>
Based on this content, identify 3-5 direct competitors for {{ company_name }}.
Only list companies that compete for the same customers.
</user>

Run multiple focused prompts and combine results in your workflow. This is what steps are for.

Writing Style Guidelines

For prompts that generate customer-facing content, define explicit style rules:

<system>
You write company research briefs for sales teams.

Writing guidelines:
- Voice: Professional but conversational, not corporate
- Objectivity: State facts, avoid superlatives ("leading", "best-in-class")
- Language: Clear and direct, no jargon unless industry-specific
- Structure: Use bullet points for lists, short paragraphs for narrative
- Length: 200-300 words unless specified otherwise
</system>

These constraints produce consistent output that matches your brand voice.

Output Shape Selection

Match the Output.* helper to your output shape:

Need	Output Helper	Example
Free-form text	(none)	Summaries, emails, explanations
Typed object	Output.object({ schema })	Extracted data, evaluator judgments
List of items	Output.array({ element })	Contacts, competitors, action items
One of N choices	Output.choice({ options })	Lead classification, sentiment

// Classification → Output.choice
const { output: intent } = await generateText({
  prompt: 'classify_lead@v1',
  variables: { activity: leadActivity },
  output: Output.choice({ options: ['high', 'medium', 'low', 'unknown'] })
});

// Extraction → Output.object
const { output: company } = await generateText({
  prompt: 'extract_company@v1',
  variables: { content: websiteContent },
  output: Output.object({ schema: companySchema })
});

// List → Output.array
const { output: competitors } = await generateText({
  prompt: 'find_competitors@v1',
  variables: { companyName: 'Acme Corp', industry: 'SaaS' },
  output: Output.array({ element: z.object({ name: z.string(), reason: z.string() }) })
});

Further Reading

• Anthropic’s prompt engineering guide
• LLM-as-a-Judge Best Practices — Writing effective judge prompts specifically