davidkim | AI

Claude Skills Are Real

So Anthropic just publicly announced their new skills feature which confirmed what Simon Willison discovered last week about how Claude was being improved for common document tasks (PowerPoint, Excel, Word, PDF). Skills are a reuseable file system that takes advantage of Claude's computer use, providing specific context when needed to appropriate topics. More information can be found in Anthropic's newly added Claude docs and cookbook, as well as their new skills repo.

This context engineering layer is a simple yet generalizable feature, especially for domain specific knowledge in enterprise settings . The progressive disclosure system is useful in ensuring scalability with context rot and I think we will see many more of these context window optimizations as agents scale in production.

I like the modularity and scalability of skills in comparison to MCPs, and could totally see this as the next step in guided tool use.

Here's some code snippets from test trial with the prompt "create a slack meme gif of a recent famous meme" (Claude chose "Chill Guy"):

cd /home/claude && cat > chill_guy_complete.py << 'EOF'
from PIL import Image, ImageDraw
import math
import imageio

WIDTH = 480
HEIGHT = 480
FPS = 15
NUM_FRAMES = 45  # 3 seconds

bg_color = (240, 248, 255)
char_color = (100, 150, 200)

def draw_text_outlined(draw, text, position, font_size, text_color, outline_color):
    """Simple text with outline"""
    x, y = position
    # Draw outline
    for offset_x in [-2, 0, 2]:
        for offset_y in [-2, 0, 2]:
            if offset_x != 0 or offset_y != 0:
                draw.text((x + offset_x, y + offset_y), text, 
                         fill=outline_color, anchor="mm")
    # Draw main text
    draw.text(position, text, fill=text_color, anchor="mm")

def draw_chill_guy(draw, x, y, size, wave_offset):
    """Draw the chill guy character"""

There's also an mcp-builder skill that I'm looking forward to experimenting with. The fact that a complex process such as MCP server development was able to be distilled and streamlined within a zip package and file framework gives me hope for the future of a capability engineering pipeline built on top of LLMs.