Learning Skills from Skills: The Next Evolution of Skillz

From Tool Building to Skill Acquisition

In my last post, I shared how Skillz allows AI to create its own executable tools at runtime. This was a massive step forward for efficiency—reducing token usage by 98% and enabling complex workflows.

But building tools from scratch isn't always the answer. Sometimes, the knowledge already exists. It's just locked away in a repository somewhere.

That's why today, I'm announcing the next major evolution of Skillz: Skill Acquisition.

---

Feature 1: Git-Based Skill Repositories

There's a growing ecosystem of "Knowledge Skills" (promoted by Anthropic's Claude Skills initiative) alongside executable tools. These are essentially markdown files (`SKILL.md`) that teach an AI expert knowledge about a specific domain.

Skillz now has native support for these repositories. With a single command, your AI can ingest an entire library of expertise.

Now, when you ask "How do I build an MCP server?", Claude seamlessly accesses the mcp-builder knowledge skill, retrieving the official best practices guide, and applies it to your request.

It supports both:

• Kepano/Obsidian style: Skills in `skills/skill-name/SKILL.md`
• ComposioHQ style: Skills in root `skill-name/SKILL.md`

---

Feature 2: Progressive Disclosure & Bundles

With great power comes great context window clutter. If you have 500 tools available, your system prompt becomes massive and expensive.

We solved this with Progressive Disclosure.

Tools are now grouped into "Skills". By default, most are hidden. The AI (or you) activates them only when needed.

The Workflow

1. Discovery: The AI uses `suggest_skill` to find relevant capabilities based on your request.
2. Activation: It enables the skill with `activate_skill`.
3. Execution: The tools appear, the job gets done.
4. Cleanup: The skill is deactivated, keeping your context clean.

We also introduced Skill Bundles. Instead of activating tools one by one, you can activate a bundle:

# Activate the full developer suite
activate_skill(name="@dev-full")
# → Activates: dev, sys, memory, and git-tools

---

Feature 3: LLM-Assisted Discovery

How does the AI know what it doesn't know? We implemented LLM-Assisted Discovery.

Using the MCP `sampling` capability, the `suggest_skill` tool effectively "asks around" in the background.

---

Putting It All Together

The synergy between these features—Git Import + Knowledge Skills + Progressive Disclosure—creates a system that feels truly intelligent.

It's no longer just a "tool user." It's a "learner." It can:

1. Realize it lacks a capability.
2. Search its library of potential skills (local or git-based).
3. Activate the right skill (executable tool or knowledge guide).
4. Solve the problem.
5. Deactivate the skill to save resources.

This brings us one step closer to the vision of Self-Evolving AI Agents.