Skills

The repository ships five skills. They describe workflows, not logic: which command to reach for, how to read the answer, and which mistakes produce a confident wrong conclusion. No business logic is duplicated in them.

Installing

zabbix-ai-cli-mcp skills list
zabbix-ai-cli-mcp skills install claude
zabbix-ai-cli-mcp skills install codex

Both runtimes read skills/<name>/SKILL.md with YAML front matter, so one set of files serves both. By default they go to the user's directory (~/.claude/skills or ~/.codex/skills); --project installs into the current directory instead.

An existing skill is left alone. --force overwrites, and says how many it replaced.

Why they exist

The commands are discoverable through zabbix-ai-cli-mcp schema, but discoverability does not convey judgement. A skill is where the judgement lives: that suppressed problems must be reported rather than filtered out; that an empty problem list is not evidence of health until maintenance and no-data have been checked; that chaining raw API calls to rebuild host investigate reproduces the API's traps along with its data.

Writing your own

Front matter needs a name and a description. The description is what the runtime matches against when deciding whether the skill applies, so it should name the words a person would actually use.

---
name: zabbix-capacity
description: Assess Zabbix host capacity trends — disk filling up, memory pressure, growth rates. Use when asked whether a host will run out of something.
---

# Zabbix: capacity

1. `zabbix-ai-cli-mcp metrics history <host> "disk" --last 7d --json`
2. Read the `summary` block rather than every point.
3. ...

Keep them short and specific. A skill that restates the help text costs context without adding judgement.

Edit on GitHub · 2026-08-26