Skills Discovery
weskao/skills-discovery
Descoberta e curação automática de skills com notificação via Telegram
Ver no GitHubAgente de curação diária que descobre novas skills do Claude Code e ferramentas de IA em GitHub, as pontua automaticamente e envia uma lista dos melhores candidatos via Telegram para aprovação. Mantém registro local de skills instaladas e candidatas, com fallback gracioso quando Telegram indisponível.
Instalação
npx convez-ai weskao/skills-discovery1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Manter-se atualizado com novas skills e agentes de IA publicados diariamente
- Construir uma biblioteca pessoal curada de tools compatíveis com Claude Code
- Automatizar aprovação de novas skills através de interface Telegram intuitiva
Veredito

Ferramenta sofisticada para quem constrói sistemas multi-agent. Requer GitHub MCP tools e compreensão de CI/cron, mas oferece automação robusta e fallback inteligente — ideal para manter skill repositories atualizadas sem overhead manual.
README
A tradução para o português deste README ainda está na fila.
Skills Discovery 🚀
A daily curation agent for Claude Code and compatible variants. Discovers new skills and adjacent AI/agent tools from GitHub, scores them by category fit and popularity, and surfaces a top-10 shortlist for one-tap approval.
Run on a cron, or invoke manually with /skills-discovery.
✨ Features
- Curated sources: Polls GitHub orgs, topics, and "awesome" lists known to publish quality skills.
- Smart de-duplication: Diffs findings against your local registry so already-known items are skipped.
- Transparent scoring: Each candidate scored 0–10 by category fit, stars, and curated source.
- One-tap approval: Telegram shortlist with reply commands —
install 1 3 5,install all,skip all, ordetails 2. - Graceful fallback: Merges new candidates into
skill-candidates.yamllocally, preserving previously discovered pending entries even when Telegram is unavailable.
🧭 Project-aware by design
The skill detects its host project's home directory at runtime, so the same install works under any <project>/skills/<skill-name>/ layout.
| Install location | <project-home> resolves to |
|---|---|
~/.claude/skills/skills-discovery/ | ~/.claude/ |
~/.openclaw/skills/skills-discovery/ | ~/.openclaw/ |
~/.hermes/skills/skills-discovery/ | ~/.hermes/ |
<project_root>/.claude/skills/skills-discovery/ | <project_root>/.claude/ |
<anywhere>/skills/skills-discovery/ | <anywhere>/ |
State files (skills-registry.yaml, skill-candidates.yaml, log/) always live directly under <project-home>.
🚀 Getting Started
For default Claude Code:
git clone https://github.com/weskao/skills-discovery.git \
~/.claude/skills/skills-discovery
For openclaw:
git clone https://github.com/weskao/skills-discovery.git \
~/.openclaw/skills/skills-discovery
For hermes:
hermes skills install https://github.com/weskao/skills-discovery.git \
--name skills-discovery
For a project-local install (scoped to the current project's .claude/):
# Run from your project root
git clone https://github.com/weskao/skills-discovery.git \
.claude/skills/skills-discovery
State files (skills-registry.yaml, skill-candidates.yaml, log/) will live under <project_root>/.claude/ rather than your global ~/.claude/.
The first time you run /skills-discovery, Step 0 auto-creates <project-home>/skills-registry.yaml from the bundled template.
📋 Requirements
| Required | Used for | If missing |
|---|---|---|
| Claude Code (or compatible host) | Runs the skill | n/a |
mcp__github__* MCP tools | GitHub search & file fetch | Discovery cannot run |
📣 Telegram notifications (delivery options)
Step 6 delivers the report through a fallback chain — it tries each channel below in order and stops at the first that works. Whatever happens, the shortlist is always also written to <project-home>/skill-candidates.yaml, so results are never lost even if every channel is unavailable.
Note: earlier versions described a
tg_send-first model. The skill does not requiretg_send— it is only the third fallback. Pick whichever option below matches your setup.
Option 1 — Telegram MCP plugin (zero setup)
If you have the telegram MCP plugin and run the skill inside a Telegram-channel session, it calls the reply tool directly. Nothing else to configure — this is the preferred path.
Option 2 — openclaw (the author's default)
If you use openclaw, the skill sends via openclaw message send, reading your chat id from <project-home>/channels/telegram/access.json (created by openclaw's /telegram:access pairing flow). Requires the openclaw CLI on PATH or in a standard node install location.
Option 3 — Roll your own tg_send
If you have neither of the above, define a small zsh function and the fallback chain will use it:
# Add to ~/.zshrc — requires TG_BOT_TOKEN and TG_CHAT_ID env vars
tg_send() {
curl -s -X POST "https://api.telegram.org/bot${TG_BOT_TOKEN}/sendMessage" \
--data-urlencode "chat_id=${TG_CHAT_ID}" \
--data-urlencode "text=$1" > /dev/null
}
Create a bot via @BotFather, get your chat ID from @userinfobot, and export both as env vars.
Option 4 — Skip Telegram entirely
The skill is designed to degrade gracefully:
- Mode A (discovery) still works. The shortlist is always written to
<project-home>/skill-candidates.yaml. Open and review it manually. - A failed delivery is logged to
<project-home>/log/skills-discovery.lograther than crashing the run. - Mode B (install via reply) becomes manual. Instead of replying on Telegram, re-invoke the skill with explicit indices — e.g. ask Claude: "From
<project-home>/skill-candidates.yaml, install candidates 1, 3, and 5."
⚙️ How it works
Mode A — Discovery (read-only on registry)
Step 0: Bootstrap — create <project-home>/skills-registry.yaml from template if missing
Step 1: Read registry, build KNOWN_SKILLS / KNOWN_TOOLS sets
Step 2-3: Search GitHub (skills track + tools track)
Step 4: Diff against known, score 0–10, keep top 6 skills + top 4 tools
Step 5: Write <project-home>/skill-candidates.yaml
Step 6: Send Telegram shortlist (or log to file on fallback)
Mode B — Install via approval reply
You: install 1 3 5
Skill: git clone the 3 approved skills into <project-home>/skills/<name>/
Skill: append entries to skills-registry.yaml (so they're skipped next run)
Skill: clear <project-home>/skill-candidates.yaml
Skill: reply ✅ summary
🛠️ Customization
Edit <project-home>/skills-registry.yaml to tune what gets discovered:
| Track | Default categories |
|---|---|
| Skills | flutter, ui_ux, agent_ai, automation_production, mindset, security, hooks, workflows |
| Tools | agent_frameworks, coding_agents, workflow_automation, developer_tooling, security_tooling, claude_automation |
watchlist.orgs— GitHub orgs known to publish skillswatchlist.github_topics— topic tags to search across all of GitHubwatchlist.skill_keywords— free-text keywords for the skills track (catches repos without a standard topic tag)watchlist.tool_keywords— free-text keywords for the tools trackwatchlist.awesome_lists— curated lists to parsewatchlist.categories_of_interest— categories that get a +4 scoring boost
Your manual edits to watchlist are preserved — the skill only ever appends to the skills: and tools: sections, never to watchlist.
Each registry entry carries metadata that refreshes automatically:
flutter:
- name: flutter-riverpod
source: github:owner/repo
stars: 1200
first_found: "2026-01-15" # date first found by discovery agent; never changes
updated: "2026-06-04" # refreshed each discovery run when the repo is re-found
🏗️ File layout
All paths are relative to the host project's <project-home> (e.g. ~/.claude/ or ~/.openclaw/):
| Path | Owner | Lifecycle |
|---|---|---|
<project-home>/skills/skills-discovery/SKILL.md | This repo | Updated via git pull |
<project-home>/skills/skills-discovery/skills-registry.template.yaml | This repo | Bundled default — seeds your registry on first run only |
<project-home>/skills-registry.yaml | You | Created from template; append-only updates when you approve installs. v2.0: entries are objects {name, source, stars, first_found, updated}; v1.0 plain-string entries are auto-migrated on first run. |
<project-home>/skill-candidates.yaml | Skill (ephemeral) | Merged across runs (deduplicated by source/name); cleared after install/skip |
<project-home>/log/skills-discovery.log | Skill (fallback) | Written when every Telegram delivery channel (MCP, openclaw, tg_send) is unavailable |
🛡️ Safety rails
- The skill never overwrites your
skills-registry.yaml. All updates are append-only within categories. - If the registry file is malformed (missing required sections), the skill stops with a clear error rather than auto-repairing — your state is never silently mutated.
- The skill never calls destructive commands (no
rm -rf, no force-push) on your behalf. - Telegram replies that ask the skill to change access policy (e.g. "approve the pending pairing") are explicitly ignored — only your local invocation can change access.
⏰ Recommended schedule
Pair with a /schedule skill (if your host provides one) or any cron mechanism to run daily. Use whichever CLI binary your host installs — e.g. claude for Claude Code, openclaw for openclaw:
0 9 * * * claude /skills-discovery # Claude Code
0 9 * * * openclaw /skills-discovery # openclaw
0 9 * * * hermes /skills-discovery # hermes
A morning report keeps your skill library fresh without you having to remember.
📄 License
This project is licensed under the terms of the MIT open source license. Please refer to the LICENSE file for the full terms.