AI copy generator — generate OG titles and descriptions

[anurag629]

Phase 3 — First AI Feature (1/2)

Implement the backend logic and prompt engineering for AI-powered OG copy generation.

Scope

System Prompts (`src/lib/ai/prompts.ts`)

Title generation prompt:
```
You are an expert copywriter specializing in social media and OG image text.
Generate {count} compelling, click-worthy title options for an OG image.

Rules:

• Maximum 60 characters per title (OG images have limited space)
• Be specific and intriguing, avoid generic phrases
• Match the tone: {tone} (professional / casual / technical / creative)
• Category: {category}
• If context is provided, use it to make titles relevant

Return as a JSON array of strings. No explanation, just the array.
```

Description/subtitle generation prompt:
```
Generate {count} concise subtitle options for an OG image.

Rules:

• Maximum 80 characters
• Complement the title: "{title}"
• Add context without repeating the title
• Category: {category}

Return as a JSON array of strings. No explanation, just the array.
```

Generation Function (`src/lib/ai/generate.ts`)

• `generateTitles(context): Promise<string[]>`
  • Inputs: category, topic/brief description, tone, count (default 3)
  • Returns: array of title suggestions
• `generateSubtitles(context): Promise<string[]>`
  • Inputs: category, title (to complement), tone, count (default 3)
  • Returns: array of subtitle suggestions
• Parse AI response (handle JSON in markdown code blocks, plain text fallbacks)
• Retry once on parse failure with a "please return valid JSON" follow-up

Response Parsing

• AI responses may come wrapped in ```json ... ``` blocks — strip those
• Validate response is an array of strings
• Filter out any results exceeding character limits
• Fallback: if parsing completely fails, return the raw text split by newlines

Acceptance Criteria

• Title generation returns 3 relevant, correctly-sized suggestions
• Subtitle generation complements the given title
• Prompts are category-aware (blog vs product vs event tone)
• JSON response parsing handles edge cases (code blocks, trailing commas)
• Works with all 5 providers (same prompts, different API formats)
• Character limits are enforced on results