Vibe Coding Part 7: The AI Tools That Actually Work (and the Ones That Don't)

The AI Tool Landscape

There are hundreds of AI tools. Most are overpriced, underpowered, or both. After months of testing, here's what actually works for building a real product — and what doesn't.

What We Use Every Day

1. NVIDIA NIM API — Free Image Generation

NVIDIA's NIM platform gives free access to flux.1-dev — a state-of-the-art image generation model. No credit card. No usage limits (within reason).

• Quality: Production-ready. Images look like professional photography.
• Speed: 5-15 seconds per image
• Cost: $0.00
• Use case: Blog hero images, product mockups, social media content

We tested 3 different API keys. All worked. The free tier is generous enough for a small e-commerce site.

2. Edge TTS — Free Voiceovers

Edge TTS uses Microsoft's speech synthesis — the same technology behind Azure Cognitive Services. But free.

• Quality: Near-human. We compared to professional voice actors — barely distinguishable.
• Languages: 30+ including Spanish, Arabic, English, Portuguese
• Cost: $0.00
• Use case: Video voiceovers, product descriptions, accessibility

The voice ar-AE-HamdanNeural (Gulf Arabic) and es-MX-DaliaNeural (Mexican Spanish) are particularly good.

3. Claude / GPT — Code Generation

The backbone of vibe coding. We use AI models for:

• Component generation (React/Next.js)
• Database schema design
• API route creation
• Bug fixing and refactoring
• Content writing (blog posts, product descriptions)

The key: be specific in your prompts. "Create a product card component" gives generic results. "Create a product card with 60-30-10 color rule, using --brand-secondary for price, --brand-accent for Add to Cart button, with Tailwind gap-6" gives production-ready code.

What Didn't Work

Wan 2.1/2.2 — Text-to-Video

We wanted to use Wan 2.1 for AI-generated video. The results look incredible — realistic motion, cinematic quality.

The problem: hardware requirements are insane.

• Wan 2.1 (1.3B): 8GB VRAM minimum
• Wan 2.2 (14B): 24GB VRAM minimum
• Our GPU: GTX 1650 (4GB VRAM)

That's not a software problem. It's physics. No amount of vibe coding can make 4GB act like 24GB.

Our workaround: Generate still images with flux.1-dev, animate with Ken Burns effect. 90% as good, 0% of the compute cost.

Local LLMs (Ollama)

We tried running LLMs locally with Ollama. The experience:

• 7B models: Usable but slow on our hardware
• 13B models: Too slow for interactive use
• 70B models: Impossible without enterprise GPU

For vibe coding, cloud APIs are better. The speed difference is 10x. The quality difference is noticeable. And the cost is minimal (or free, in NVIDIA's case).

The Free Tier Strategy

Here's how we built a production site for nearly nothing:

• Image generation: NVIDIA NIM (free) → flux.1-dev
• Voiceovers: Edge TTS (free) → 30+ languages
• Code generation: Claude/GPT (paid, but worth it)
• Hosting: Self-hosted on existing hardware
• CDN: Cloudflare free tier
• Database: PostgreSQL (free, open source)

Total infrastructure cost: ~$20/month (domain + Cloudflare Pro). Everything else is free or self-hosted.

How to Choose AI Tools

1. Test the free tier first. Don't pay until you've exhausted free options.
2. Check hardware requirements. If a model needs 24GB VRAM and you have 4GB, move on.
3. Measure quality, not features. A tool with 10 features that works is better than 100 features that don't.
4. Consider the workflow. The best tool is the one that fits your existing process.
5. Document what works. Your future self won't remember which API key worked.

The AI tool landscape changes fast. What matters is not which tool you pick today — it's that you pick one and start building.