# Skillier

> A single AI skill that searches a curated index of 4,000 expert skills, surfaces the best matches for any task, and loads them into your AI's context in one click.

Skillier is a routing skill for AI clients like Claude Desktop, Claude Code, and OpenClaw that gives your AI on-demand access to a curated library of expert skills. Install it once, make a request as usual, and Skillier searches its index, surfaces the top 3–5 matches, and lets you activate the ones you want — loading their instructions directly into context. The project is currently in free early access, with a Lite edition that is MIT-licensed and fully open source.

## What It Is

Skillier sits between you and your AI client as a single installed skill that acts as a skill router. Rather than writing custom prompts or maintaining a library of instructions yourself, you activate Skillier and it handles discovery: it matches your natural-language request against a curated databank of expert skill cards spanning finance, design, DevOps, legal, writing, ML, and more. The matched skills are presented as candidates — you confirm which ones load, and the AI follows their instructions for the rest of the turn. Nothing activates automatically; every load requires your explicit confirmation via `AskUserQuestion`.

## Two Editions: Lite and Full

Skillier ships in two distinct editions:

- **Skillier Lite** — the open-source, local-only edition. It bundles 1,852 curated skills in a ~7 MB `.skill` file, uses pure stdlib BM25 (Robertson-Sparck-Jones) for ranking, makes no network calls, and has no telemetry. It is MIT licensed and free forever.
- **Skillier (Full)** — the hosted edition with a backend at `api.skillier.ai`. It adds BGE-small semantic vectors fused with BM25 via Reciprocal Rank Fusion, an opt-in cross-encoder reranker (Xenova/ms-marco-MiniLM-L-6-v2), and telemetry-driven pick boosting. The full client automatically falls back to the local Lite BM25 index when the backend is unreachable. The full edition's source is not open; it is distributed as a prebuilt bundle.

## How the Skill Index Is Built

The Lite databank indexes 1,852 skills across 18 upstream open-source repositories, including sources from Anthropic, Trail of Bits, Vercel Labs, and community contributors. After deduplication (same name + description across sources), cross-source fork filtering, and a curation gate, the result is packed into a single `skillbank.jsonl` file with a precomputed BM25 inverted index. Each entry carries a `source` field for attribution. The full hosted index extends this to approximately 4,000 skills with semantic search layered on top.

## Security Model

Because skill bodies become in-band instructions for the AI, Skillier treats all skill content as untrusted. The ingest pipeline strips invisible Unicode characters — including Tag block codepoints (U+E0000–E007F), zero-width characters, bidirectional overrides, and variation selectors — that could be used to hide prompt-injection instructions from a human reviewer. A `check_invisible.py` script can be wired as a pre-commit hook. The sanitizer and its test suite are open source in the repository. Low-substance and near-duplicate skills are filtered out at curation time to keep the index signal-dense.

## Update: Skillier v1.5.0

The latest release is v1.5.0, published on 2026-05-30. The GitHub repository was created in May 2026 and shows active development, with the most recent push on 2026-05-30. The release renames bundle assets — current releases use `skillier-lite.*` for the Lite edition and `skillier.*` for the full edition, reversing the naming convention used in v1.4.0 and earlier. The project is described as being in free early access for the hosted full edition.

## Features
- Routes AI requests to the best-matching expert skills from a curated index
- 4,000 expert skills in the full hosted edition; 1,852 in the open-source Lite edition
- BM25 search (Lite) with semantic BGE-small vectors and Reciprocal Rank Fusion (Full)
- Optional cross-encoder reranker (Xenova/ms-marco-MiniLM-L-6-v2) via ?rerank=true
- Telemetry-driven pick boosting for popular query→skill pairs (Full only)
- Explicit user confirmation via AskUserQuestion before any skill loads
- Automatic fallback to local BM25 index when hosted backend is unreachable
- Prompt-injection defense: strips invisible Unicode, bidi overrides, and tag-block characters
- Local extras support via local_extras.jsonl for custom skill additions
- ~7 MB .skill bundle with no native dependencies (Lite)
- No network calls, no telemetry in Lite edition
- Installs in ~30 seconds via one-liner or drag-drop into Claude Desktop/claude.com
- Supports Claude Desktop, Claude Code CLI, and OpenClaw

## Integrations
Claude Desktop, Claude Code, OpenClaw, claude.com

## Platforms
WEB, CLI, API

## Pricing
Open Source

## Version
v1.5.0

## Links
- Website: https://www.skillier.ai
- Documentation: https://github.com/skillier-ai/skillier
- Repository: https://github.com/skillier-ai/skillier
- EveryDev.ai: https://www.everydev.ai/tools/skillier