ai-content-generation（AI 内容生成）

AI Content Generation - Integration Guide

Purpose

Comprehensive guide for AI content generation in the AIProDaily platform, covering OpenAI/Claude integration, prompt management, structured outputs, and content scoring systems.

When to Use

Automatically activates when:

• Calling AI APIs (OpenAI, Claude)
• Using callAIWithPrompt()
• Creating or modifying prompts
• Working with app_settings AI prompts
• Implementing content generation
• Building scoring systems
• Handling AI responses

Core Pattern: callAIWithPrompt()

Standard Usage

Location: src/lib/openai.ts

import { callAIWithPrompt } from '@/lib/openai'

// Generate article title
const result = await callAIWithPrompt(
  'ai_prompt_primary_article_title',  // Prompt key in app_settings
  newsletterId,                        // Tenant context
  {
    // Variables for placeholder replacement
    title: post.title,
    description: post.description,
    content: post.full_article_text
  }
)

// result = { headline: "AI-Generated Title" }

How It Works

1. Loads prompt from app_settings table by key + newsletter_id
2. Replaces placeholders like {{title}}, {{content}} with provided variables
3. Calls AI API (OpenAI or Claude) with complete request
4. Parses response according to response_format schema
5. Returns structured JSON object

Key Features

✅ Database-driven: All prompts stored in database, not hardcoded ✅ Tenant-scoped: Each newsletter can customize prompts ✅ Type-safe: JSON schema enforces response structure ✅ Flexible: Supports both OpenAI and Claude ✅ Reusable: Same function for all AI operations

Prompt Storage Format

Database Schema

-- app_settings table
CREATE TABLE app_settings (
  key TEXT PRIMARY KEY,
  value JSONB NOT NULL,
  description TEXT,
  newsletter_id UUID NOT NULL,
  ai_provider TEXT,  -- 'openai' or 'claude'
  created_at TIMESTAMPTZ DEFAULT NOW(),
  updated_at TIMESTAMPTZ DEFAULT NOW()
)

Complete Prompt Structure

INSERT INTO app_settings (key, value, newsletter_id, ai_provider, description)
VALUES (
  'ai_prompt_primary_article_title',
  '{
    "model": "gpt-4o",
    "temperature": 0.7,
    "max_output_tokens": 500,
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "article_title_response",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "headline": {
              "type": "string",
              "description": "The generated article headline"
            }
          },
          "required": ["headline"],
          "additionalProperties": false
        }
      }
    },
    "messages": [
      {
        "role": "system",
        "content": "You are an expert headline writer for accounting professionals..."
      },
      {
        "role": "user",
        "content": "Source Title: {{title}}\n\nSource Content: {{content}}\n\nWrite a compelling headline."
      }
    ]
  }'::jsonb,
  'newsletter-uuid-here',
  'openai',
  'Content Generation - Primary Article Title: Generates engaging headlines'
);

All parameters stored in database:

• model - AI model to use
• temperature - Creativity level (0-1)
• max_output_tokens - Response length limit
• response_format - JSON schema for structured output
• messages - System and user prompts with placeholders

Response Format Patterns

Simple String Response

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "simple_response",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "result": { "type": "string" }
        },
        "required": ["result"],
        "additionalProperties": false
      }
    }
  }
}

Complex Structured Response

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "article_body_response",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "headline": { "type": "string" },
          "body": { "type": "string" },
          "summary": { "type": "string" },
          "key_points": {
            "type": "array",
            "items": { "type": "string" }
          }
        },
        "required": ["headline", "body"],
        "additionalProperties": false
      }
    }
  }
}

Scoring Response

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "content_score_response",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "score": {
            "type": "number",
            "minimum": 0,
            "maximum": 10
          },
          "reasoning": { "type": "string" }
        },
        "required": ["score", "reasoning"],
        "additionalProperties": false
      }
    }
  }
}

Multi-Criteria Scoring System

Overview

Purpose: Evaluate RSS posts using multiple weighted criteria Location: src/lib/rss-processor.ts Storage: post_ratings table

Configuration

// Criteria settings in app_settings
{
  "criteria_enabled_count": 3,  // 1-5 criteria
  "criteria_1_name": "Interest Level",
  "criteria_1_weight": 1.5,
  "criteria_2_name": "Relevance",
  "criteria_2_weight": 1.5,
  "criteria_3_name": "Impact",
  "criteria_3_weight": 1.0
}

Scoring Process

// Each criterion gets separate AI call
for (let i = 1; i <= criteriaCount; i++) {
  const promptKey = `ai_prompt_criteria_${i}`
  const weight = settings[`criteria_${i}_weight`]

  // Call AI for this criterion
  const result = await callAIWithPrompt(
    promptKey,
    newsletterId,
    {
      title: post.title,
      description: post.description,
      content: post.content
    }
  )

  // Store individual score
  await supabaseAdmin
    .from('post_ratings')
    .insert({
      post_id: post.id,
      newsletter_id: newsletterId,
      criterion_name: criteriaName,
      score: result.score,        // 0-10
      weighted_score: result.score * weight,
      reasoning: result.reasoning
    })
}

// Calculate total score (sum of weighted scores)
const totalScore = ratings.reduce((sum, r) => sum + r.weighted_score, 0)

Example Scoring

Criterion 1: Interest Level (weight 1.5) → score 8 → weighted 12.0
Criterion 2: Relevance (weight 1.5)     → score 7 → weighted 10.5
Criterion 3: Impact (weight 1.0)        → score 6 → weighted 6.0
═══════════════════════════════════════════════════════════════
Total Score: 28.5

Prompt Design Best Practices

System Message

{
  "role": "system",
  "content": `You are an expert content writer for accounting professionals.
Your audience is CPAs, accountants, and financial professionals.
Write in a professional yet engaging tone.
Focus on practical, actionable information.
Keep content concise and scannable.`
}

User Message with Placeholders

{
  "role": "user",
  "content": `Source Article:
Title: {{title}}
Description: {{description}}
Full Content: {{content}}

Task: Write a 200-300 word article summary that:
1. Captures the key takeaways
2. Explains why this matters to accountants
3. Uses clear, professional language
4. Ends with a thought-provoking statement

Output the summary as a JSON object with a "body" field.`
}

Temperature Guidelines

// Creative content (headlines, summaries)
"temperature": 0.7

// Factual content (analysis, scoring)
"temperature": 0.3

// Consistent output (classifications)
"temperature": 0.1

Model Selection

OpenAI Models

// Fast, cost-effective (most common)
"model": "gpt-4o"

// Latest, most capable
"model": "gpt-4o-2024-11-20"

// Smaller, faster for simple tasks
"model": "gpt-4o-mini"

Claude Models

// Most capable
"model": "claude-3-5-sonnet-20241022"

// Fast, cost-effective
"model": "claude-3-5-haiku-20241022"

// Older, still powerful
"model": "claude-3-opus-20240229"

Error Handling

Standard Pattern

try {
  const result = await callAIWithPrompt(
    promptKey,
    newsletterId,
    variables
  )

  // Validate response
  if (!result || !result.headline) {
    throw new Error('Invalid AI response: missing required fields')
  }

  return result

} catch (error: any) {
  console.error('[AI] Error calling AI:', error.message)

  // Check for specific errors
  if (error.message.includes('rate_limit')) {
    console.error('[AI] Rate limit exceeded, implement backoff')
  }
  if (error.message.includes('context_length')) {
    console.error('[AI] Input too long, need to truncate')
  }

  throw error
}

Retry with Backoff

async function callAIWithRetry(
  promptKey: string,
  newsletterId: string,
  variables: Record<string, any>,
  maxRetries = 2
) {
  let retryCount = 0

  while (retryCount <= maxRetries) {
    try {
      return await callAIWithPrompt(promptKey, newsletterId, variables)
    } catch (error: any) {
      retryCount++

      // Don't retry on validation errors
      if (error.message.includes('Invalid')) {
        throw error
      }

      if (retryCount > maxRetries) {
        throw error
      }

      console.log(`[AI] Retry ${retryCount}/${maxRetries} after error`)
      await new Promise(resolve => setTimeout(resolve, 2000 * retryCount))
    }
  }
}

Rate Limiting

OpenAI Limits

Tier 1 (free/trial):

• gpt-4o: 500 requests/day
• gpt-4o-mini: 10,000 requests/day

Tier 2 (paid):

• gpt-4o: 5,000 requests/min
• gpt-4o-mini: 30,000 requests/min

Batching Strategy

// Process in batches to avoid rate limits
const BATCH_SIZE = 3
const BATCH_DELAY = 2000  // 2 seconds between batches

const batches = chunkArray(posts, BATCH_SIZE)

for (const batch of batches) {
  // Process batch in parallel
  await Promise.all(
    batch.map(post => generateArticle(post))
  )

  // Wait before next batch
  if (batches.indexOf(batch) < batches.length - 1) {
    await new Promise(resolve => setTimeout(resolve, BATCH_DELAY))
  }
}

console.log(`[AI] Processed ${posts.length} items in ${batches.length} batches`)

Content Generation Workflows

Article Title Generation

const titleResult = await callAIWithPrompt(
  'ai_prompt_primary_article_title',
  newsletterId,
  {
    title: rssPost.title,
    description: rssPost.description,
    content: rssPost.full_article_text
  }
)

// Store generated title
await supabaseAdmin
  .from('articles')
  .insert({
    newsletter_id: newsletterId,
    campaign_id: campaignId,
    rss_post_id: rssPost.id,
    headline: titleResult.headline,
    article_text: null  // Body generated separately
  })

Article Body Generation

const bodyResult = await callAIWithPrompt(
  'ai_prompt_primary_article_body',
  newsletterId,
  {
    title: rssPost.title,
    headline: article.headline,  // Use AI-generated headline
    description: rssPost.description,
    content: rssPost.full_article_text
  }
)

// Update with generated body
await supabaseAdmin
  .from('articles')
  .update({
    article_text: bodyResult.body
  })
  .eq('id', article.id)
  .eq('newsletter_id', newsletterId)

Fact-Checking

const factCheckResult = await callAIWithPrompt(
  'ai_prompt_fact_check',
  newsletterId,
  {
    headline: article.headline,
    body: article.article_text,
    source_content: article.rss_post.full_article_text
  }
)

// Store fact-check score
await supabaseAdmin
  .from('articles')
  .update({
    fact_check_score: factCheckResult.score,
    fact_check_reasoning: factCheckResult.reasoning
  })
  .eq('id', article.id)
  .eq('newsletter_id', newsletterId)

Testing Prompts

Test in Isolation

// Create test route: app/api/test/prompt/route.ts
export async function POST(request: NextRequest) {
  const { promptKey, variables } = await request.json()

  try {
    const result = await callAIWithPrompt(
      promptKey,
      'test-newsletter-id',
      variables
    )

    return NextResponse.json({
      success: true,
      result
    })
  } catch (error: any) {
    return NextResponse.json({
      error: error.message
    }, { status: 500 })
  }
}

export const maxDuration = 60

Validate Response Schema

function validateArticleResponse(result: any): boolean {
  if (!result) return false
  if (typeof result.headline !== 'string') return false
  if (typeof result.body !== 'string') return false
  if (result.headline.length < 10) return false
  if (result.body.length < 50) return false
  return true
}

Best Practices

✅ DO:

• Store all prompts in app_settings database
• Use JSON schema for response format validation
• Include clear instructions in system message
• Use placeholders for dynamic content
• Implement retry logic for transient errors
• Batch API calls to respect rate limits
• Validate AI responses before using
• Log AI calls for debugging
• Use appropriate temperature for task
• Test prompts thoroughly before production

❌ DON'T:

• Hardcode prompts in code
• Skip response validation
• Ignore rate limits
• Use overly complex prompts
• Forget error handling
• Expose API keys client-side
• Use wrong model for task
• Trust AI output blindly
• Skip testing with real data
• Make unbatched parallel calls

Troubleshooting

AI Returns Invalid Format

Check:

1. JSON schema is correct
2. strict: true is set
3. Instructions are clear
4. Model supports structured outputs

Rate Limit Errors

Solutions:

1. Implement batching (3-5 requests per batch)
2. Add delays between batches (2-5 seconds)
3. Use retry with exponential backoff
4. Upgrade API tier if needed

Content Quality Issues

Improve:

1. Refine system message instructions
2. Adjust temperature (lower for consistency)
3. Provide better examples in prompt
4. Add validation rules
5. Use more capable model

Timeout Errors

Fix:

1. Reduce max_output_tokens
2. Simplify prompt
3. Use faster model (gpt-4o-mini, claude-haiku)
4. Increase API route maxDuration

Reference

Main Function: src/lib/openai.ts - callAIWithPrompt() Prompt Storage: app_settings table Response Storage: articles, post_ratings tables Scoring Logic: src/lib/rss-processor.ts

Related Docs:

• docs/AI_PROMPT_SYSTEM_GUIDE.md
• docs/OPENAI_RESPONSES_API_GUIDE.md
• docs/workflows/MULTI_CRITERIA_SCORING_GUIDE.md

Skill Status: ACTIVE ✅ Line Count: < 500 ✅ Integration: OpenAI + Claude ✅