Week 2 Theme: Tooling for AI Tools

The shift

We have spent years building tools for humans: buttons, dashboards, command lines, forms, and workflows that assume a person is staring at the screen. Week 2 asks everyone to flip the perspective. What if the primary user is an AI agent - an API-first, context-hungry, occasionally stateless coworker?

There are roughly eight billion humans. There are going to be a lot more software agents. So this theme is simple: build tools for the future users of software.

What to build

This week is about making something AI-native. Not just AI-powered. Not a wrapper around a model. A useful tool that an AI system can actually reach for, call, parse, and use to move work forward.

Your submission can be a standalone tool that produces game art, an API integration, a Claude/Codex/Copilot skill, an OpenClaw tool, a context hydrator, or anything else that gives an AI system a new capability. Copilot and Codex count as simple AI systems. OpenClaw counts too. It is a deliberately tolerant theme.

Start with Agent Skills

The things you can do with AI skills are jaw-dropping. Take five minutes with AgentSkills.io; the return on that read can be a real unlock. The pattern is straightforward: describe when the skill should activate, give the agent the right interface, and return structured outputs that can be used in the next step.

Starter ideas

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

Texture Rescuer

Upload a degraded image and get back a crisp, PBR-ready asset: tiled, named, and ready for Godot or Unity.

Context Hydrator

Feed in a URL; get scraped, summarized, structured context back into the conversation for the agent to use.

Style Unpacker

Drop in a reference image and return a structured prompt plus parameters that reproduce the aesthetic.

Sequence Stitcher

Let the model plan a multi-shot video sequence while your tool handles the ffmpeg assembly.

Agent Skill

Build any capability you wish existed: image to code, code to diagram, noise to music, file to checklist.

API Test Rig

Give an agent a safe tool for probing APIs, saving fixtures, and returning structured failure reports.

The process matters

The first time I built an AI skill was a material unlock in what I could do. If you are wondering what simple thing you can do right now that would change what you can do with AI, tooling for AI tools is high on the list.

Where humans tend to get tripped up is the process. I did. So I recorded a small example of using AI tools to build a tool that AI uses.

Video summary

In the walkthrough, I introduce a new YouTube channel and use it to bridge fast-moving AI concepts with real work. The concrete demo is building an Agent Skill in Visual Studio Code: the skill.md file, the scripts folder, and the workflow of giving an AI assistant a capability it can invoke.

The example tries to create an RPG character generator skill with a Python script, hits environment trouble and an MCP server conflict, then converts the script to PowerShell with AI help. There are laughs, learnings, and a few tears with Python - which is also a useful picture of what the process actually feels like.

Design notes for agent-facing tools

Prefer structured output. JSON beats pretty tables when the next reader is an agent.
Make the activation obvious. A good skill description helps the model know when to use the tool.
Keep the interface narrow. Agents do better with clear inputs, explicit flags, and predictable returns.
Assume iteration. Tool-building with AI is still software work: test it, debug it, simplify it, then document it.

Source context: This first-party Spring into AI theme page is adapted from my original Advisory Hour Substack post, Week 2 Theme: Tooling for AI Tools. For future competition updates, you can also subscribe to Advisory Hour.

No matter which path you take, there is one thing for certain.

Week 2 has begun. Good luck.