Week 2 • February 17–23, 2026

🔧 Tools for AI Tools

Build something designed for AI to use. When the user is an agent, not a human, everything changes.

The Shift

We've spent years building tools for humans—buttons, dashboards, CLI commands. But the user is changing. Now it's AI agents doing the work, and most tools weren't designed for an API-first, context-hungry, stateless coworker.

This week: Build something AI-native. Not a wrapper around GPT. Not just "AI-powered." A tool designed for AI to use—where the primary interface is function calls, context windows, and structured output.

🎵 See It In Action: Git Jukebox

Real submission from this week's competition. A complete walkthrough of building an AI skill that generates music from git commits — built entirely via Discord while away from the computer.

The 6-step workflow — from brainstorming to packaging and deployment
Progressive disclosure explained — how skills load efficiently
Links to the code — GitHub repo with full source
The Substack write-up — full backstory with audio examples

View the Workflow Guide →

💡 What If... (Starter Ideas)

Use these as sparks, not requirements. Remix, mutate, or ignore entirely.

🛠️ Texture Rescuer

Upload a degraded image, get back a crisp PBR-ready asset—auto-tiled, named, and ready for import into Godot/Unity. The model calls it when it hits "needs better graphics" mid-project.

📚 Context Hydrator

Feed a URL to your tool, it scrapes, summarizes, and injects structured data back into the conversation. AI-powered research that feeds itself.

🎨 Style Unpacker

Drop in a reference image, get back a structured prompt + parameter set that reproduces that aesthetic. The agent reverse-engineers vibes.

🎬 Sequence Stitcher

Multi-shot video assembly where the model plans the sequence, your tool handles the ffmpeg stitching. Prompt-to-cut.

🔧 Claude Skill

Any capability you wish existed—image→code, code→diagram, noise→music. If an agent could reach for one tool, what should it be?

🆚 VS Code: Building AI Skills

VS Code extensions can expose commands that AI agents invoke. Here's the pattern:

// package.json - register AI-callable commands
{
  "contributes": {
    "commands": [
      { "command": "aiSkill.analyzeCode", "title": "AI: Analyze Code" },
      { "command": "aiSkill.generateDiagram", "title": "AI: Generate Diagram" }
    ]
  }
}
            

// Your command handler - designed for AI calls
export async function analyzeCode(
  uri: string,
  options: { depth: 'shallow' | 'deep'; focus?: string[]; }
): Promise<AnalysisResult> {
  // Return structured JSON, not just UI updates
  return {
    complexity: calculateComplexity(uri),
    dependencies: extractDependencies(uri),
    suggestions: generateSuggestions(uri, options),
    exportable: true // Flag for AI consumption
  };
}
            

Pro Tip: Design your VS Code extension outputs as structured JSON that AI can parse and act on. Return analysis results, not just pretty-printed messages.

Current VS Code AI Extensions (2025)

GitHub Copilot — Inline suggestions, context-aware completions
RooCode / KiloCode — Agent-mode coding assistants
Gemini Code Assist — Google's AI pair programmer
Agent Skills Marketplace — Discover and install AI capabilities
Continue.dev — Open-source AI code assistant

🤖 OpenAI Codex: Skills & MCP

Codex CLI now supports Skills and Model Context Protocol (MCP)—two ways to extend what AI can do.

Option A: Codex Skills (Beginner-Friendly)

Skills are self-contained capabilities in ~/.codex/skills/. Codex matches them to your task.