[FEATURE] Skill Search/Retrieval Tool

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Organizations building agents quickly accumulate thousands of reusable skills (playbooks, workflows, domain-specific procedures, etc). A naïve approach—sending the full skill catalog to the LLM/agent at runtime—doesn’t scale and becomes counterproductive.

Core problem

As the number of skills grows, agents need a retrieval/search layer to determine which small subset of skills to expose to the model for a given user task and context. Without this layer, skill usage becomes unreliable, expensive, and hard to govern.

Why “return everything to the LLM” is inefficient

• Context window constraints: A large skill library cannot fit in the prompt. Even if it could, it would crowd out the actual user/task context.
• Decision quality degrades: Too many options increases confusion and mis-selection; the model may pick a suboptimal or unsafe skill.
• Maintainability issues: Skill catalogs change constantly (versions, deprecations, new owners/tags). Prompt-based routing becomes brittle.

What’s needed

A first-class Skill Retrieval/Search mechanism that, given a task + context, returns a ranked set of relevant skills (multiple candidates) with the right metadata/constraints—so the agent can choose and execute the best option while keeping prompts small, fast, and safe.

Proposed Solution

Introduce a first-class Skill Search retrieval layer for Claude agents, modeled after the existing Tool Search pattern. Rather than returning an organization’s entire catalog of thousands of skills to the model, the agent should be able to call a skill_search interface with the current task + context and receive a ranked shortlist of multiple relevant skills (not just one).

Like Tool Search, this should support custom implementation so organizations can plug in their own skill registry and return the best candidates with lightweight metadata (name, description, etc). This keeps prompts small, lowers cost/latency, improves skill selection quality, and enables safer governance as skill libraries scale.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Developer tools/SDK

Use Case Example

_No response_

Additional Context

_No response_