DEV Community

정상록
정상록

Posted on

Google Agent Skills: How a Single SKILL.md File Boosted AI Coding Accuracy from 6.8% to 96%

The Knowledge Gap Problem

If you've used AI coding agents, you've hit this wall: you ask for code using a new SDK, and the agent generates deprecated API calls. Google quantified this as the "Knowledge Gap" — the mismatch between static LLM training data and rapidly evolving software ecosystems.

Their benchmark puts it starkly: Gemini 3.0 Pro scored just 6.8% on Gemini SDK code generation without skill assistance. That means 93 out of 100 code generation attempts failed.

What Are Agent Skills?

An Agent Skill is a standalone instruction package that bridges this gap:

my-skill/
  SKILL.md           # Core instructions for the agent
  references/        # Official docs excerpts
  assets/            # Diagrams, images
  scripts/           # Helper scripts
Enter fullscreen mode Exit fullscreen mode

The SKILL.md file loads into the agent's context window, providing up-to-date instructions on how to use specific APIs and frameworks correctly.

The Benchmark Results

Google tested gemini-api-dev skill with 117 prompts across Python and TypeScript:

Model Vanilla With Skill Improvement
Gemini 3.0 Pro 6.8% 96% 14x
Gemini 3.0 Flash 6.8% 87% 13x
ADK-specific 29% 99% 3.4x

One file. 14x improvement. That's not a marginal gain — it's a paradigm shift in how we should be using AI coding agents.

ADK Integration

Google's Agent Development Kit makes skill integration straightforward:

from google.adk.skills import SkillToolset
from google.adk.agents import Agent

skill = SkillToolset.load_skill_from_dir("./skills/gemini-api-dev")

agent = Agent(
    name="coding_agent",
    model="gemini-3.0-pro",
    tools=[skill],
    instruction="You are a Gemini API coding agent."
)
Enter fullscreen mode Exit fullscreen mode

Cross-Platform Standard

Here's what makes this particularly interesting: SKILL.md is becoming a de facto industry standard. The same skill file works across:

  • Claude Code (Anthropic)
  • Cursor (AI-native IDE)
  • Windsurf (AI pair programming)
  • Gemini CLI (Google)
  • Codex CLI (OpenAI)
  • GitHub Copilot

The awesome-agent-skills repo has 1,234+ skills registered, with management tools like skillport and openskills emerging.

Building Your Own Skill

Creating a skill is surprisingly simple:

  1. Write SKILL.md — Instructions for the agent (what API to use, what's deprecated)
  2. Add references/ — Excerpted official docs, migration guides, code examples
  3. Deploy — Place in .claude/skills/, .cursor/skills/, or load via ADK's SkillToolset

The key is being specific about what's correct (latest API patterns) and what's forbidden (deprecated patterns).

Key Takeaway

The knowledge gap is a structural limitation of LLMs — no amount of model scaling fixes it. Agent Skills provide a practical engineering solution. Google's 6.8% to 96% benchmark proves the approach works at scale.

Check out the gemini-skills repo and the Google Developers Blog post for details.

What frameworks are you building Agent Skills for? I'd love to hear about community experiences with SKILL.md across different AI coding tools.

Top comments (0)