# Gemini Blog

> Configure or debug LLM blog post generation using Vercel AI SDK and Google Gemini. Use when updating blog generation prompts, fixing AI integration issues, modifying content generation logic, or working with structured output schemas.

- Skill: `motormetrics/gemini-blog-2` (Agent Skill)
- Install (CLI): `npx skillmds@latest add motormetrics/gemini-blog-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/motormetrics/gemini-blog-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: motormetrics (https://skillmd.com/u/motormetrics)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/motormetrics/gemini-blog-2

---


# Gemini Blog Generation Skill

Blog generation package: `packages/ai/`

## Architecture

```
packages/ai/
├── src/
│   ├── generate-post.ts    # 2-step generation (analysis → structured output)
│   ├── config.ts           # System instructions
│   ├── schemas.ts          # Zod schemas (postSchema, highlightSchema)
│   ├── tags.ts             # Tag constants (CARS_TAGS, COE_TAGS)
│   ├── hero-images.ts      # Hero image URLs
│   └── save-post.ts        # Post persistence with idempotency
```

### 2-Step Flow

1. **Step 1 (Analysis)**: `generateText()` + Code Execution Tool + Extended Thinking → Accurate calculations
2. **Step 2 (Generation)**: `generateObject()` + Zod schema → Type-safe structured output

## Key Functions

```typescript
// Generate and persist (returns saved-post metadata, not the raw Zod object)
import { generateBlogContent } from "@motormetrics/ai";

const post = await generateBlogContent({
  data: tokenisedData,     // Pipe-delimited data
  month: "October 2024",
  dataType: "cars",        // "cars" or "coe"
});

// post.postId, post.title, post.slug, post.excerpt
```

## Schemas

```typescript
// postSchema
z.object({
  title: z.string().max(100),              // SEO title, max 60 chars preferred
  excerpt: z.string().max(500),            // Meta description, under 300 chars
  content: z.string(),                     // Markdown (no H1)
  tags: z.array(z.string()).min(1).max(10), // 3-5 tags, first is dataType
  highlights: z.array(highlightSchema),    // 3-6 key statistics
});

// highlightSchema
z.object({
  value: z.string(),   // "52.60%", "$125,000"
  label: z.string(),   // "Electric Vehicles Lead"
  detail: z.string(),  // "2,081 units registered"
});
```

## Tag Constants

```typescript
export const CARS_TAGS = ["Cars", "Registrations", "Fuel Types", "Market Trends", ...] as const;
export const COE_TAGS = ["COE", "Quota Premium", "1st Bidding Round", "PQP", ...] as const;
```

## Updating Prompts

Edit `packages/ai/src/config.ts`:
- `ANALYSIS_INSTRUCTIONS`: For calculation logic
- `GENERATION_INSTRUCTIONS`: For output format

## Debugging

**Low Quality Output:** Check Step 1 analysis logs, verify Code Execution Tool runs Python
**Schema Validation Errors:** Check Zod constraints (max lengths, array bounds)
**API Errors:** Verify `AI_GATEWAY_API_KEY`, check Gateway quota/billing

## Environment Variables

```env
AI_GATEWAY_API_KEY=...              # Required (Vercel AI Gateway)
DATABASE_URL=...                    # Required for generate-and-save (Neon/Postgres)
BLOB_READ_WRITE_TOKEN=...           # Required for local hero-image upload
LANGFUSE_PUBLIC_KEY=pk-lf-...       # Optional telemetry
LANGFUSE_SECRET_KEY=sk-lf-...
```

## Best Practices

1. **Always use 2-step flow**: Separate analysis from generation
2. **Never skip Code Execution**: Required for accurate calculations
3. **Use tag constants**: Maintain vocabulary consistency
4. **Enable telemetry**: Track costs and quality

## References

- `packages/ai/CLAUDE.md` for full package documentation
- Vercel AI SDK: Use Context7 for latest docs