scitix/skill-authoring

Skill Authoring Guide

Read this guide before creating a new skill or improving an existing one.

Directory Layout

A skill is a directory package, not a single file. The smallest valid skill is:

<skill-name>/
└── SKILL.md

Optional package files can live beside SKILL.md:

<skill-name>/
├── SKILL.md
├── scripts/        # executable .sh/.py helpers
├── references/     # markdown docs the agent can read
├── examples/       # example inputs/outputs
└── assets/         # small images or other package assets

The directory name must match the name field in SKILL.md. Use uppercase SKILL.md; lowercase skill.md is not canonical.

SKILL.md Format

---
name: <kebab-case-name>
description: >-
  One-line summary. Mention the execution tool if the skill uses scripts.
---

Followed by markdown body:

# <Title>

## Purpose
What problem this skill solves and when to use it.

## Tool
<execution tool invocation — required for script-based skills>
Example: local_script: skill="check-pod-oom", script="check.sh", args="<ns> <pod>"

## Parameters
| Parameter | Required | Description |
|-----------|----------|-------------|
| `<arg1>`  | Yes      | ...         |

## Procedure
Step-by-step actions with concrete commands.

## Examples
Concrete tool invocations with realistic parameters.

Script Execution Modes

| Tool | Runs where | When to use | |------|-----------|-------------| | local_script | AgentBox (local) | kubectl commands from outside the cluster — most common | | node_script | K8s node (host) | Needs host tools, /proc, /sys, devices, nsenter | | pod_script | Inside a pod | Diagnostics inside a running container | | node_script + netns | Node + pod's network ns | Host tools + pod's network view (call resolve_pod_netns first) |

Best Practices

• One skill, one purpose — don't make Swiss army knives
• Script over inline commands — if the procedure has more than 3 steps, write a script
• Idempotent and read-only — skills are for diagnosis, not remediation
• Description is critical — the agent uses it to decide whether to use this skill
• Concrete examples — at least 2 examples with realistic parameters
• Explain expected output — describe what normal vs abnormal output looks like
• Severity thresholds — when checking error counters, give thresholds so the agent can judge

Common Mistakes

• Too much raw output — dump kubectl describe without filtering. Grep for relevant lines
• Missing Tool section — without it, the agent doesn't know which execution tool to use
• Wrong execution mode — using local_script for host-level tools (use node_script)
• Hardcoded values — node names, namespaces should be parameters
• No severity guidance — e.g., CRC errors: 0 = normal, 1-100 = minor, >1000 = critical