Skillier

A single Claude skill that searches a packed databank of 1,852 curated expert skills with stdlib BM25, surfaces the top candidates, lets you multi-select via AskUserQuestion, and loads the chosen SKILL.md bodies so Claude follows them.

This repository is the open-source home of Skillier Lite — the local-only, no-network, no-dependency, no-telemetry edition. The source you see here is the product: clone it, read it, run it.

The repo's Releases are also the canonical download point for prebuilt bundles of both editions:

Bundle	Edition	Open source?
skillier-lite.skill / skillier-lite.tar.gz	Skillier Lite — same thing as this repo, packaged for drag-drop install	✅ Yes (this repo)
skillier.skill / skillier.tar.gz	Skillier — the full hybrid-semantic client with the hosted backend baked in	❌ No — distributed as a build only

Asset naming: current releases use the names above (skillier-lite.* for Lite, skillier.* for Skillier). Older releases (≤ v1.4.0) used skillier.* for Lite and skillier-full.* for the full edition.

Skillier (the full edition — hosted at skillier.ai, backend at https://api.skillier.ai) adds BGE-small semantic search fused with BM25 via Reciprocal Rank Fusion, an opt-in cross-encoder reranker, telemetry-driven pick boost, and an admin stats endpoint. Its runtime build is downloadable from Releases here, but its source is not open. Lite stays the offline fallback even when you install Skillier — the full client falls back to this same local BM25 index automatically when the backend is unreachable.

---

How it works

1. Claude calls scripts/search.py "<query>" → JSON array of candidate skills with scores.
2. Claude shows the candidates to you via AskUserQuestion (multi-select).
3. For each pick, Claude calls scripts/load.py <id> → returns the full SKILL.md body prefixed with a "Skill loaded" preamble.
4. Claude follows the loaded skill's instructions for the remainder of the turn.

SKILL.md in this directory is the contract Claude reads. Ranking (BM25) and storage (packed JSONL + precomputed inverted-index JSON) are implementation details Claude doesn't need to know about.

Install

Claude Code CLI

git clone https://github.com/skillier-ai/skillier ~/.claude/skills/skillier

Claude auto-discovers the skill on the next session start.

Claude Desktop or claude.com (drag-drop)

1. Open the latest release.
2. Download skillier-lite.skill (Lite) or skillier.skill (Skillier — backend URL baked in).
3. Open Claude Desktop OR go to claude.com → sidebar → Customize → Skills → click + → Create skill → Upload a skill → drop the file.

Or build the Lite bundle yourself from this repo: python3 scripts/pack.py writes ~/Desktop/skillier-lite.zip (and a .skill copy). pack.py respects Claude.app's import limits (≤1024-char description, ≤200 files, ≤30 MB unzipped), which is why the databank ships as a single packed skillbank.jsonl instead of thousands of .md files.

OpenClaw

git clone https://github.com/skillier-ai/skillier ~/.openclaw/workspace/skills/skillier

One-liner (any host)

curl -fsSL https://skillier.ai/install | sh                # Lite (local-only, free, MIT)
curl -fsSL https://skillier.ai/install | sh -s -- --full   # Skillier (hosted backend)

The install script detects your host and downloads the matching bundle from this repo's latest release.

Sources

The databank is seeded from upstream skill repositories (see scripts/seed.py). Every skillbank.jsonl row carries a source field with the upstream repo URL for attribution. After dedup (same name + description across sources), cross-source fork filtering, the curation gate (see scripts/curate.py), and dropping two auto-generated bulk sources to keep the bundle under 30 MB unzipped: 1,852 skills indexed across 18 sources.

Source	Skills
davila7/claude-code-templates	661
alirezarezvani/claude-skills	386
secondsky/claude-skills	203
wshobson/agents	149
K-Dense-AI/claude-scientific-skills	139
trailofbits/skills	74
garrytan/gstack	56
bmad-code-org/BMAD-METHOD	41
obra/superpowers-skills	31
ComposioHQ/awesome-claude-skills	22
mhattingpete/claude-skills-marketplace	18
anthropics/skills	17
obra/superpowers	14
google-gemini/gemini-cli	13
google-labs-code/stitch-skills	13
vercel-labs/agent-skills	6
obra/superpowers-lab	5
anthropics/claude-cookbooks	4
Total	1,852

Adding your own skills

Local additions live in local_extras.jsonl — one JSON row per skill, committed alongside skillbank.jsonl. pack.py calls merge_local_extras at build time, sanitizing each row through the security defense layer and deduping against upstream entries.

Use the helper for ergonomics:

scripts/add_local_skill.py path/to/SKILL.md         # single skill
scripts/add_local_skill.py path/to/dir/             # recursive directory walk
scripts/add_local_skill.py path/to/bundle.skill --strip-frontmatter

Idempotent — re-running with the same id replaces the row.

Security

Skill content reaches the LLM verbatim, so a malicious skill author could embed invisible instructions via Unicode Tag block (U+E0000–E007F) or zero-width / bidi / variation-selector chars. The ingest pipeline strips every codepoint in INVISIBLE_RANGES (see scripts/sanitize.py) and logs each detection. scripts/check_invisible.py can be wired as a pre-commit hook to block injected SKILL.md at PR time. Both the sanitizer and its test suite (scripts/tests/test_sanitize.py) are open in this repo.

Repo layout

SKILL.md                  # frontmatter triggers + flow instructions Claude reads
skillbank.jsonl           # packed skill databank (one skill per line)
local_extras.jsonl        # hand-curated additions (committed; merged by pack.py)
index/bm25.json           # precomputed BM25 postings + meta
scripts/
├── search.py             # stdlib BM25 query, prints JSON — Claude calls this
├── load.py               # prints a skill's body by id — Claude calls this
├── sanitize.py           # invisible-char stripper (defense layer, see Security)
├── check_invisible.py    # pre-commit hook script for skillbank content
├── add_local_skill.py    # helper: parse SKILL.md/.skill bundle → append to local_extras.jsonl
├── build_index.py        # rebuilds index/bm25.json from skillbank.jsonl or skillbank/
├── curate.py             # quality/curation gate over the raw seeded corpus
├── seed.py               # clones upstream repos → skillbank/*.md
└── pack.py               # md_to_jsonl + merge_local_extras → skillbank.jsonl + .zip/.skill

skillbank/ (raw per-file form from upstream repos) is gitignored — regenerated via scripts/seed.py.

Refresh the databank

scripts/seed.py --clean   # re-clones all upstream repos → skillbank/*.md
scripts/pack.py           # writes skillbank.jsonl, rebuilds index/bm25.json, emits ~/Desktop/skillier-lite.zip

Edit SOURCES in scripts/seed.py to add upstream repos (and add the matching name → URL entry to SOURCE_URLS in scripts/pack.py so attribution stays honest). Edit SKIP_PREFIXES in scripts/pack.py to adjust the low-value filter.

License

MIT. See LICENSE.

The 1,852 skills bundled in skillbank.jsonl remain the property of their original authors under the licenses of the upstream source repositories. Each entry carries a source field with the upstream URL for attribution.

Contact

Issues, takedown requests, questions: hi@skillier.ai or open a GitHub issue.