Add reusable skills

Some instructions are too long to keep in the system prompt but too important to leave out. Your incident triage procedure, the severity matrix, the report template - an agent needs all of it when an incident comes in, and none of it the rest of the time.

Skills solve that. Each one packages instructions plus supporting files, and the agent loads the detail only when the situation calls for it. What sits in the prompt permanently is a short catalog entry; the body arrives on demand.

Create the skill beneath axl-config/skills/:

axl-config/skills/incident-triage/
├── SKILL.md
├── severity-matrix.md
└── report-template.md

SKILL.md carries the metadata and the instructions:

---
name: incident-triage
description: Triage incidents using the team severity matrix
tags: [incident, operations]
resources:
  - severity-matrix.md
  - report-template.md
---

# Instructions

Classify impact before recommending a response.

Then enable it on the agent:

skills = ["incident-triage"]

Writing one that gets used

The description is the part that decides whether a skill ever fires, because it is always visible to the agent while the instructions are not. "Triage incidents using the team severity matrix" tells the model when to reach for this. "Incident stuff" does not.

Keep SKILL.md focused on procedure and push bulky reference material into resources. A skill whose instructions are themselves enormous defeats the purpose - you have just moved the prompt bloat somewhere else.

If your deployment requires signature verification, sign the skill and add its author to the configured keyring. See Verify configuration content.

Skills or RAG?

Both hold information the agent does not need every time, and the distinction is about shape rather than size.

Whether to classify a Sev-2 is a skill question. What the Q3 incident report said about latency is a RAG question.

Next

• Add RAG - for knowledge that is searched rather than followed.
• Verify configuration content - signing skills and enforcing a policy.
• Agent configuration reference - the complete schema.