Stop "Prompt Engineering". Start "Skill Searching"

#programming #ai #huggingface #python

Prompt Engineering is dead. Long live Skill Searching.

We’ve all been there: tweaking a prompt for hours ("You are a helpful assistant...", "Take a deep breath...") trying to get a model to behave. It’s fragile, unscalable, and frankly, boring.

Enter UPskill by Hugging Face (conceptually). The idea is simple: instead of writing prompts, you generate them using a "Teacher" model, validate them against test cases, and save the winner as a reusable "Skill".

I built a Local-Only Demo to prove this works without sending a single token to the cloud. Here’s how I did it using Ollama and Gemma 3 (12B).

The Problem: Vague Inputs

Users are lazy. They type things like:

"write something about AI"

Baseline models will give you a generic, rambling encyclopedia entry. But what if you wanted a structured JSON object for your app?

The Solution: A "Prompt Optimizer" Skill

I created a meta-skill that intercepts these vague requests and transforms them into structured, high-quality prompts (or in this case, pure JSON for validation).

Step 1: The "Teacher" (Gemma 3)

I asked Gemma 3 to generate a system prompt that forces JSON compliance.

{
  "role": "Tech Journalist",
  "goal": "Explain AI basics",
  "constraints": "Max 200 words",
  "output_format": "Markdown"
}

Step 2: Verification

I ran this skill against a test suite of vague inputs.

Baseline (Llama/Gemma raw): Fails 100% of the time (returns text).
Skilled Model: Passes 100% of the time (returns valid JSON).

For developers, this means deterministic behavior from non-deterministic models. You can treat "Skills" like library functions—import them, trust them, and stop worrying about the underlying prompt magic.