What if you could teach Claude Code everything you know about your infrastructure? Not generic advice — like "how to deploy a Laravel app" — but exactly how YOU deploy on YOUR server, with YOUR paths, YOUR services, YOUR SLOs.
That's exactly what I did. And I open-sourced the whole thing.
The problem: context that evaporates
Claude Code is absurdly capable. But it knows nothing about your server. Every time you open a new session, you need to explain again: "my Nginx config lives here", "my MySQL has these settings", "my deploy follows these steps".
You end up repeating the same context, session after session. It's like having a brilliant SRE who loses their memory every morning.
I run a production server with 5 Laravel apps, 3 Next.js apps, MySQL, Redis, Supervisor, PM2, Grafana, Loki — the full stack. I manage everything through a Telegram bot integrated with Claude Code. And I realized the bottleneck wasn't the AI's capability — it was context.
The solution: Skills — runbooks for AI
Skills are structured instructions that get injected into Claude's context when relevant. Think of them as runbooks, but written to be consumed by an AI, not by a human reading documentation at 3 AM during an incident.
Each skill follows Anthropic's official plugin marketplace format:
plugins/devops/sre/
├── .claude-plugin/
│ └── plugin.json
└── skills/sre/
└── SKILL.mdThe plugin.json defines metadata:
{
"name": "SRE Server Operations",
"description": "Full SRE runbook: health checks, SLOs, incident response...",
"version": "1.0.0"
}And SKILL.md is where the real knowledge lives — with frontmatter, context, commands, runbooks, and best practices. Pure Markdown, easy to write, easy to maintain.
How automatic selection works
This is where it gets interesting. When someone messages my Telegram bot — "check if MySQL is slow" — the system needs to decide which skill to inject into context.
The approach? TF-IDF. Classic NLP. No embeddings, no API calls, zero cost.
The system tokenizes the user query, calculates similarity against each skill's description, and injects the most relevant ones. It works surprisingly well for this use case — skill descriptions are short and specific enough that TF-IDF handles it just fine.
Query: "MySQL is slow, check the queries" → Injected skills: mysql-performance, sre
Simple, fast, no external dependencies.
34 skills across 7 categories
I organized everything into categories that reflect the daily work of anyone managing infrastructure:
DevOps (9 skills)
The core. From a complete SRE runbook with SLOs, health checks, and incident response, to zero-downtime deploys for Laravel and Next.js. There are skills for API troubleshooting, Nginx load balancing, performance profiling, cron scheduling, Linux service diagnostics, and Cloudflare integration.
The sre skill alone is a complete manual: daily checklists, runbooks for every incident type (site down, MySQL down, Redis full, disk full, high CPU), useful Loki queries, and deploy procedures.
Development (7 skills)
Laravel debugging, PHP/Laravel full-stack development, advanced Git operations, and my personal favorite: unfuck-my-git-state. Seriously, how many times have you ended up with a detached HEAD, phantom worktrees, or a rebase gone wrong? This skill fixes it.
There's also Playwright for E2E testing and GitHub CLI for PR/issue operations straight from the terminal.
Database (4 skills)
mysql-performance is the one I use most. EXPLAIN analysis, slow query digest, index optimization — all parameterized for my environment. There's also backup/restore with Cloudflare R2, Redis optimized for Laravel, and SQL schema design.
Monitoring (3 skills)
Grafana alerting, LogQL queries for Loki, and systemd journal log streaming. Covers the entire observability stack.
Marketing (4 skills)
Yes, marketing. meta-ads-advanced covers 2026 Meta Ads strategies with Andromeda AI, server-side CAPI, Advantage+. There's growth hacking, digital marketing fundamentals, and even viral TikTok/Reels scriptwriting.
Productivity (5 skills)
agent-memory is special — persistent semantic memory across sessions. Claude forgets everything when the session ends, but with this skill it knows where and how to persist important context. There's also Google Workspace, image generation via Gemini, and web research.
Security (2 skills)
SSH and SSL/TLS. SSH command reference and certificate management with Let's Encrypt automation.
The power of specificity
The difference between a generic AI and an AI that knows your environment is massive.
Generic: "To deploy a Laravel application, run git pull, composer install, php artisan migrate..."
With your server's skill:
# Zero-downtime deploy - myapp.example.com
cd /home/deploy/myapp.example.com
sudo -u deploy php artisan down --retry=60
sudo -u deploy git pull origin main
sudo -u deploy composer install --no-dev --optimize-autoloader
sudo -u deploy php artisan migrate --force
sudo -u deploy php artisan config:cache
sudo -u deploy php artisan route:cache
sudo -u deploy php artisan view:cache
sudo -u deploy php artisan queue:restart
sudo -u deploy php artisan up
curl -I https://myapp.example.com
The second one knows the correct user (deploy, not root), knows which caches to clear, knows to verify the health check afterward. It's the difference between generic advice and a tested operational procedure.
Auto-sync: update on GitHub, all devices get it
A simple cron handles distribution:
# Every 30 minutes, sync skills
*/30 * * * * cd /path/to/claude-skills && git pull --quietEdited a skill on GitHub? Within 30 minutes, every device using the repo has the updated version. No deploy, no CI/CD, no fuss.
How to create your own skill
It's simple. Create the directory structure:
mkdir -p plugins/your-category/your-skill/.claude-plugin
mkdir -p plugins/your-category/your-skill/skills/your-skillWrite the plugin.json:
{
"name": "My Skill",
"description": "What this skill does in one sentence",
"version": "1.0.0"
}And the SKILL.md with the knowledge:
---
name: my-skill
description: Description for semantic matching
---
# My Skill
## Context
When and why to use this skill.
## Commands
Commands specific to your environment.
## Runbooks
Step-by-step procedures.
## Best Practices
What works, what doesn't.
Done. Next time someone asks something related, the system will inject this skill automatically.
Open source
The repository is public: github.com/billyfranklim1/claude-skills
There are 34 ready-to-use skills. You can install the entire collection via Anthropic's marketplace:
/plugin marketplace add billyfranklim1/claude-skillsOr clone the repo and adapt it to your environment. Fork it, change the paths, change the services, change the SLOs. The format is the same.
If you manage infrastructure with Claude Code — or want to start — these skills are a solid starting point. And if you want to contribute new skills, PRs are welcome.
At the end of the day, the idea is simple: don't repeat context. Teach once, use always.
