I Built a 48-Skill Arsenal for My AI Agent — Here's What's Inside

The Problem with Generic AI Agents

Most AI agents are great at being chatty. Not so great at being useful.

Out of the box, an AI agent can answer questions. What it cannot do is:

• Review a pull request and comment with specific line-level feedback

• Run a semantic search across your entire codebase

• Set up a private search engine for your homelab in 10 minutes

• Wake up each morning and brief you on what happened overnight

• Improve itself after every failure

I needed an agent that could actually do work — not just talk about work.

That is what openclaw-arsenals is.

---

What Is It?

A production-ready skill and agent collection for OpenClaw, built over 6 months of real use.

48 skills, 13 named agents, 16 infrastructure guides, and 10 automation workflows — all sanitized, documented, and tested.

Clone it:

git clone https://github.com/adelpro/openclaw-arsenals.git
cp -r * ~/.openclaw/skills/

---

Skills

Write and Research

• humanizer — strips AI patterns from generated text. Makes outputs sound human, not robotic

• purple-cow-content — applies the Seth Godin / MrBeast virality lens to any content task

• reddit-readonly — curated Reddit research without API keys

• search-cluster — Google, Wikipedia, Reddit, and RSS in one query

• answeroverflow — searches Discord community discussions that never make it to Google

Code Quality

• pr-review — automated PR review with blast-radius analysis

• skill-engineer — builds skills with Designer / Reviewer / Tester quality gates

• git-workflow — conventional commits, branch strategy, PR etiquette

• github-ops — full repo lifecycle: create → push → release

• github-actions — CI/CD workflow generation with security hardening

Agent Orchestration

• council-of-wisdom — 3+ agents deliberate before a decision is made

• agent-orchestrator — decomposes a task, spawns sub-agents, consolidates results

• agent-team-orchestration — multi-agent teams with defined roles and handoff protocols

Self-Improvement (The Compounding Loop)

foundry is a self-writing agent. It observes failures, crystallizes patterns into permanent hooks, and evolves its own tools. Run it every night and your agent wakes up smarter than when it went to sleep.

The loop:

• Work → Foundry records outcome

• Failure repeated 5x → Hook crystallized (permanent fix)

• Success → Pattern stored for next time

• Nightly → compound-engineering reviews sessions, updates memory

Infrastructure

• private-web-search-searchxng — self-hosted search, zero tracking

• qdrant-ollama — semantic code search with local embeddings

• context7 — documentation with citations for 100+ libraries

• guardian — AI-powered penetration testing automation

• docker-essentials — container management, debugging, compose patterns

---

Agents

Named agents with defined roles:

• ORION — task orchestration and routing

• ATLAS — operations, health checks, morning briefings

• LENS — code quality, PR reviews, blast-radius analysis

• PROBE — API endpoint health and uptime alerts

• TRACE — bug analysis and regression detection

• ECHO — content, README, blog posts

• RANK — SEO audits and traffic analysis

• SCROLL — Reddit and community research

• NEWSLETTER — weekly digest curation

• VULN-SCANNER — dependency audits, CVE checks, secret scanning

---

Infrastructure Guides

Step-by-step setup guides for:

• Karakeep (bookmark storage + AI retrieval)

• SearXNG (private search)

• Qdrant + Ollama (semantic search)

• Cloudflare Tunnel (expose homelab securely)

• Portainer (container management UI)

• GitHub MCP (GitHub as MCP server)

• SocratiCode (semantic code search)

• Secrets management (zero secrets in workspace)

• Heartbeat monitoring (lightweight agent health checks)

• Documents parsing (PDF → Markdown for RAG pipelines)

---

Workflows

Ten end-to-end pipelines:

• second-brain-karakeep — save link → auto-tag → summarise → store

• pr-review-pipeline — PR → security scan → blast-radius → comment

• daily-digest — GitHub + weather + news → morning briefing

• code-to-release — branch → CI → merge → release → announce

• semantic-code-search — natural language → file references

• content-publishing-pipeline — research → write → humanize → visual → publish

• skill-creation-pipeline — idea → spec → build → review → publish

• homelab-health-check — Docker health → alert → auto-restart

• multi-agent-task-pipeline — ORION decompose → parallel agents → consolidate

• seo-audit-workflow — crawl → fix → rank → report

---

Security

Every skill in this repo has been sanitized. No API keys. No internal hostnames. No personal URLs.

Workspace files reference key names only — never the actual values. Secrets live in systemd environment variables, not in markdown files.

Before any commit, this runs:

grep -rE "ak2_|ghp_|eyJ|sk-|=[a-zA-Z0-9]{20,}"   --include="*.md" . 2>/dev/null

---

The Compounding Effect

Most automation breaks down over time. You build it, it works, then the world changes and it silently stops working.

openclaw-arsenals is designed to be self-maintaining. The foundry loop means failures generate permanent fixes. The memory stack means context compounds. The heartbeat monitoring means infrastructure drift is caught automatically.

After 6 months of use, the agent is measurably better at its job than it was on day one.

---

Next steps:

• Star the repo: https://github.com/adelpro/openclaw-arsenals

• Clone it and try one workflow

• Open an issue with a skill request