Convert Prompts to .prompt Files

Overview

This skill guides the conversion of Flow SDK inline prompts, XML prompts, and JavaScript prompt arrays to Output SDK .prompt files with YAML frontmatter.

When to Use This Skill

During Migration:

• Converting prompts.ts or prompts.xml to .prompt files
• Extracting inline prompts from activities to separate files
• Setting up prompt versioning

Flow SDK Prompt Formats

Flow SDK uses several prompt formats that need conversion:

1. Inline Prompts in Activities

// activities.ts
const prompt = `You are an assistant. Analyze: ${text}`;
const response = await completion( { messages: [ { role: 'user', content: prompt } ] } );

2. JavaScript Prompt Arrays (prompts.ts)

// prompts.ts
export const analyzePrompt = [
  { role: 'system', content: 'You are an expert analyst.' },
  { role: 'user', content: 'Analyze this: {{text}}' }
];

3. XML Prompts (prompts.xml)

<prompt name="analyze">
  <system>You are an expert analyst.</system>
  <user>Analyze this: {{text}}</user>
</prompt>

Output SDK .prompt File Format

Basic Structure

---
provider: openai
model: gpt-4o
temperature: 0.7
---

<system>
System message here.
</system>

<user>
User message with {{ variable }} interpolation.
</user>

YAML Frontmatter Fields

Field	Type	Required	Description
provider	string	Yes	openai or anthropic
model	string	Yes	Model identifier
temperature	number	No	0-1 sampling temperature
max_tokens	number	No	Maximum output tokens

Common Model Values

OpenAI:

• gpt-4o
• gpt-4-turbo
• gpt-3.5-turbo

Anthropic:

• claude-3-5-sonnet-20241022
• claude-3-opus-20240229
• claude-3-haiku-20240307

Conversion Process

Step 1: Identify All Prompts

Find prompts in the Flow SDK workflow:

# Check for prompt files
ls src/workflows/my-workflow/prompts.*

# Check for inline prompts in activities
grep -n "role: 'system'" src/workflows/my-workflow/activities.ts
grep -n "role: 'user'" src/workflows/my-workflow/activities.ts

Step 2: Create .prompt File

Name format: promptName@version.prompt

analyzeDocument@v1.prompt
generateSummary@v1.prompt
extractEntities@v1.prompt

Step 3: Convert Content

From Inline Prompt:

// Before (activities.ts)
const systemPrompt = 'You are a document analyzer.';
const userPrompt = `Analyze this document: ${documentText}`;

# After (analyzeDocument@v1.prompt)
---
provider: openai
model: gpt-4o
temperature: 0.3
---

<system>
You are a document analyzer.
</system>

<user>
Analyze this document: {{ documentText }}
</user>

From JavaScript Array:

// Before (prompts.ts)
export const summarizePrompt = [
  { role: 'system', content: 'You summarize text concisely.' },
  { role: 'user', content: 'Summarize: {{text}}\nMax length: {{maxLength}}' }
];

# After (summarize@v1.prompt)
---
provider: openai
model: gpt-4o
temperature: 0.5
---

<system>
You summarize text concisely.
</system>

<user>
Summarize: {{ text }}
Max length: {{ maxLength }}
</user>

From XML:

<!-- Before (prompts.xml) -->
<prompt name="extract">
  <system>You extract key entities from text.</system>
  <user>
    Extract entities from:
    {{#if includeContext}}
    Context: {{context}}
    {{/if}}
    Text: {{text}}
  </user>
</prompt>

# After (extract@v1.prompt)
---
provider: openai
model: gpt-4o
temperature: 0.2
---

<system>
You extract key entities from text.
</system>

<user>
Extract entities from:
{% if includeContext %}
Context: {{ context }}
{% endif %}
Text: {{ text }}
</user>

Step 4: Update Step to Use Prompt File

// Before (activities.ts)
import { summarizePrompt } from './prompts';

export async function summarize( text: string ): Promise<string> {
  const response = await completion( {
    model: 'gpt-4',
    messages: summarizePrompt.map( m => ( {
      ...m,
      content: m.content.replace( '{{text}}', text )
    } ) )
  } );
  return response.content;
}

// After (steps.ts)
import { step, z } from '@outputai/core';
import { generateText } from '@outputai/llm';

export const summarize = step( {
  name: 'summarize',
  inputSchema: z.object( { text: z.string() } ),
  outputSchema: z.string(),
  fn: async ( input ) => {
    const { result } = await generateText( {
      prompt: 'summarize@v1',
      variables: {
        text: input.text
      }
    } );
    return result;
  }
} );

Template Syntax Conversion

Important: Convert Handlebars to Liquid.js syntax!

Handlebars	Liquid.js
{{variable}}	{{ variable }}
{{#if cond}}	{% if cond %}
{{/if}}	{% endif %}
{{#each items}}	{% for item in items %}
{{/each}}	{% endfor %}
{{else}}	{% else %}

See flow-convert-handlebars-to-liquid for detailed conversion rules.

Complete Migration Example

Before: prompts.ts (Flow SDK)

export const analyzeDocumentPrompt = [
  {
    role: 'system',
    content: `You are a document analysis expert. Analyze documents for:
- Key themes
- Important entities
- Sentiment
- Action items`
  },
  {
    role: 'user',
    content: `Document Type: {{documentType}}

{{#if previousAnalysis}}
Previous Analysis:
{{previousAnalysis}}
{{/if}}

Document Content:
{{content}}

Provide a comprehensive analysis.`
  }
];

After: analyzeDocument@v1.prompt (Output SDK)

---
provider: openai
model: gpt-4o
temperature: 0.3
max_tokens: 4000
---

<system>
You are a document analysis expert. Analyze documents for:
- Key themes
- Important entities
- Sentiment
- Action items
</system>

<user>
Document Type: {{ documentType }}

{% if previousAnalysis %}
Previous Analysis:
{{ previousAnalysis }}
{% endif %}

Document Content:
{{ content }}

Provide a comprehensive analysis.
</user>

After: steps.ts (Using the Prompt)

import { step, z } from '@outputai/core';
import { generateText, Output } from '@outputai/llm';
import { AnalysisResultSchema, AnalysisResult } from './types.js';

const AnalyzeDocumentInputSchema = z.object( {
  documentType: z.string(),
  content: z.string(),
  previousAnalysis: z.string().optional()
} );

export const analyzeDocument = step( {
  name: 'analyzeDocument',
  inputSchema: AnalyzeDocumentInputSchema,
  outputSchema: AnalysisResultSchema,
  fn: async ( input ) => {
    const { output } = await generateText( {
      prompt: 'analyzeDocument@v1',
      variables: {
        documentType: input.documentType,
        content: input.content,
        previousAnalysis: input.previousAnalysis || ''
      },
      output: Output.object( {
        schema: AnalysisResultSchema
      } )
    } );

    return output;
  }
} );

Prompt Naming Convention

{descriptiveName}@{version}.prompt

Examples:
- analyzeDocument@v1.prompt
- generateSummary@v1.prompt
- extractEntities@v2.prompt
- translateContent@v1.prompt

Verification Checklist

• All prompts extracted to .prompt files
• YAML frontmatter includes provider and model
• Template syntax converted to Liquid.js
• Variable spacing correct: {{ var }} not {{var}}
• Steps use generateText() with prompt reference (use Output.object() for structured output)
• Prompt file names follow naming convention

Related Skills

• flow-convert-handlebars-to-liquid - Template syntax conversion
• flow-convert-activities-to-steps - Step conversion
• flow-analyze-prompts - Prompt cataloging