A Claude skill that auto-discovers, installs, and activates the right skills for any task — zero manual steps.
Works in Claude Code, Claude.ai (Cowork), and any Claude interface that supports skills — on Windows, macOS, and Linux.
Type "зроби сам", "do it automatically", or "set up everything" — and Claude handles the rest.
✓ Skills already present: docx, pdf-reading, xlsx
✓ Installed new: graphify 0.17.1 → hooks registered, graph built
✓ Task completed: extracted PDF data into Q2-sales.xlsx
No manual steps. No "please install X". Claude does it.
- Python 3.8+ on PATH (used by the skill scanner —
python3orpython) - Claude Code (any OS) or Claude.ai with Cowork / Skills enabled
- Optional, only if a task needs them:
uvorpipx(Python CLIs),npm(Node CLIs),curl
No other setup. The skill installs whatever a given task requires.
Option A — .skill file (Claude.ai)
Download auto-skill-manager.skill → Settings → Skills → Install from file
Option B — Claude Code (macOS / Linux)
mkdir -p ~/.claude/skills/auto-skill-manager/scripts
base=https://raw.githubusercontent.com/VovikP/auto-skill-manager/main
curl -fsSL $base/SKILL.md > ~/.claude/skills/auto-skill-manager/SKILL.md
curl -fsSL $base/scripts/scan_skills.py > ~/.claude/skills/auto-skill-manager/scripts/scan_skills.pyOption C — Claude Code (Windows / PowerShell)
$d = "$env:USERPROFILE\.claude\skills\auto-skill-manager"
New-Item -ItemType Directory -Force "$d\scripts" | Out-Null
$base = "https://raw.githubusercontent.com/VovikP/auto-skill-manager/main"
curl.exe -fsSL "$base/SKILL.md" -o "$d\SKILL.md"
curl.exe -fsSL "$base/scripts/scan_skills.py" -o "$d\scripts\scan_skills.py"Say any of these and the skill activates automatically:
| Language | Phrase |
|---|---|
| English | "do it automatically", "set up everything", "install what's needed", "run without manual steps" |
| Ukrainian | "зроби сам", "автоматично", "без ручної роботи" |
| Any | Starting a new project, opening a codebase, asking to configure tools |
If unsure whether to trigger — it triggers anyway. Safe to run on every task.
The skill runs a 7-step pipeline:
Runs scripts/scan_skills.py — a cross-platform, recursive scanner that checks every skill root for the current environment and OS, not just the Linux sandbox:
- Claude.ai / Cowork:
/mnt/skills/{public,user,examples} - Claude Code (Win/macOS/Linux):
~/.claude/skills,~/.claude/plugins - Project-local:
./.claude/skills - Windows agent-mode sessions:
%APPDATA%/Claude/local-agent-mode-sessions
python scripts/scan_skills.py # name | description per skill
python scripts/scan_skills.py --json # machine-readableIt reads each skill's YAML frontmatter and builds a live index — always reflects what's actually installed, never guesses. (The earlier flat /mnt/skills-only scan returned zero skills on Windows/macOS; this fixes that.)
Identifies file types, domain, tools, and operation type from the user's request.
| Task | Skill loaded |
|---|---|
| Word document | docx |
| PDF create/fill | pdf |
| PDF read/extract | pdf-reading |
| Spreadsheet | xlsx |
| Presentation | pptx |
| UI / frontend | frontend-design |
| Uploaded file | file-reading |
| Human-sounding text | humanizer |
| Codebase navigation | graphify (external) |
Multiple skills loaded simultaneously when needed.
If nothing in the index covers the requirement:
- GitHub →
claude code skill <topic> SKILL.md site:github.com - PyPI → web search →
uv tool install - npm →
npm install -g - Fallback → informs user, proceeds without a dedicated skill
# Claude Code skill (cross-platform path)
mkdir -p ~/.claude/skills/<name>/ # Cowork sandbox: /mnt/skills/user/<name>/
curl -fsSL <raw_github_url>/SKILL.md > ~/.claude/skills/<name>/SKILL.md
# External CLI tool
uv tool install <package> # preferred — auto-manages PATH
pipx install <package> # fallback
<tool> claude install # register with Claude Code (if supported)
<tool> hook install # git hooks for auto-rebuild (if supported)Reads the full SKILL.md, follows its workflow, completes the task. Zero hand-off to the user.
graphify hook install # rebuild on git commit
graphify global add .graphify/graph.json --as <project> # older builds: graphify-out/
<tool> install --project # writes to .claude/, commits with repoEvery run ends with a 3–5 line summary:
✓ Skills already present: docx, pdf-reading, xlsx
✓ Installed new: graphify 0.17.1 → hooks registered, graph built
✓ Task completed: [what was done]
- Never asks the user to run a command if Claude can run it directly
- Never installs from unverified sources (GitHub public repos + official PyPI only)
- Always checks before installing — no duplicate installs
- Always reads a skill after installing — never assumes behavior from the name
- Loads multiple skills simultaneously when needed
MIT