on-call-skill-creator

name	on-call-skill-creator
description	Create a new on-call assistant skill for a rotation or team, modelled on the Merchant Center one. Use this whenever a team wants their own on-call skill, or someone says "create an on-call skill", "scaffold on-call for <team>", "adopt the Merchant Center on-call skill for us", or "make on-call-<rotation>". It interviews you, ingests your runbooks, dashboards, logs, and escalation setup, and generates a ready `on-call-<rotation>` skill. Trigger even when the user only names their rotation and asks for on-call help tooling.
disable-model-invocation	false
argument-hint	<rotation-name>
allowed-tools	Read, Write, Edit, Bash
scope	["knowledge","monitoring"]

On-Call Skill Creator

Scaffold a new on-call-<rotation> skill by interviewing the owner and ingesting their sources. The reference implementation is on-call-merchant-center. This skill generalises it so any rotation can adopt the same shape.

You produce a new skill under skills/on-call-<rotation>/ with a filled SKILL.md, a README.md, and resources/. Then you register it, lint it, and hand triggering off to skill-creator.

The idea: invariant vs per-rotation

Most of an on-call skill is the same across rotations. The workflow, the safety rules, the incident process, the LogScale language, the observability stack. That part lives in resources/templates/ and resources/shared/ and you copy it.

What changes per rotation is a small set of inputs. Runbook indexes, dashboards, environments, the log package, service repos, escalation, and a persona. You gather those in the interview and by auto-discovery, then fill the placeholders.

Ask only for what you cannot discover.

Required

Same stack as the generated skills. The atlassian MCP (Confluence), the agatha-mcp Grafana MCP, the incident-io MCP, and authenticated gh. This skill also needs Write and Edit to generate files. It runs read only gh and MCP reads. It never runs operational commands.

Process

1. Parse the rotation

Take the rotation name from $ARGUMENTS. Derive the skill name on-call-<rotation> in kebab case. Confirm it with the user.

2. Interview

Read resources/interview.md and run the interview. Collect the per-rotation inputs. Rotation title, persona name and vibe, runbook index pages, services and repos, standing dashboards, the log package, escalation defaults, knowledge docs, the rotation rules page, alert phrasings, the scope tag, and 3 to 5 real incidents for validation. Offer defaults. Skip anything you can auto-discover in step 3.

3. Auto-discover

Read resources/ingestion-playbook.md. Reduce the interview by discovering:

• Environments. grafana_list_datasources(type=prometheus) builds the env to cloud id table.
• Runbook taxonomy. Read the index descendants and detect Diagnostic: and Action: naming.
• Service to repo. Prefer the Github repo field in runbook metadata.

4. Ingest sources

Follow the playbook for each input:

• Runbooks. getConfluencePageDescendants on the index pages. Metadata only.
• Dashboards. grafana_get_dashboard_panel_queries for the signal to panel map.
• Escalation. The incident.io Service catalog Escalation Path attribute.
• Knowledge docs. Fetch child pages and distill each into a topic resource with a source link.

5. Assemble the skill

Create skills/on-call-<rotation>/. Fill and copy each template into the new skill. SKILL.md.template and README.md.template go to the root. The rest go into the new skill's resources/.

• resources/templates/SKILL.md.template → the new SKILL.md
• resources/templates/README.md.template → the new README.md
• resources/templates/runbook-sources.md.template → runbook-sources.md
• resources/templates/grafana-metrics.md.template → grafana-metrics.md
• resources/templates/logscale-queries.md.template → logscale-queries.md
• resources/templates/glossary.md.template → glossary.md
• resources/templates/escalation-paths.md.template → escalation-paths.md
• resources/templates/rotation-and-rules.md.template → rotation-and-rules.md

Copy verbatim:

• resources/shared/incident-process.md → incident-process.md
• resources/shared/logscale-syntax.md → logscale-syntax.md

Generate architecture-*.md files from the ingested knowledge docs, one topic per file, each with its source link. Build an architecture-overview.md that indexes every resource. Replace every {{PLACEHOLDER}}. Leave none behind.

6. Fill the soul

The SKILL.md.template carries the persona section. Fill {{PERSONA}} and {{PERSONA_VIBE}}. Keep the safety rules verbatim. Every generated skill is a guide not an operator. It suggests kubectl, helm, and vault. It never runs them.

7. Register and lint

Add the new skill to the root README.md table and to skills/README.md. Run ./scripts/lint-skills.sh skills/on-call-<rotation>. Fix any error.

8. Validate

Spot check against the 3 to 5 real incidents from the interview. For each, confirm the skill would reach the right runbook and query. Report the result honestly. Do not claim a match you did not verify.

9. Hand off

Offer to run skill-creator for the description triggering loop. Offer to open a PR. Do not push to main.

Placeholders

Fill every one. Source in brackets.

Placeholder	Meaning
{{ROTATION}}	kebab name, e.g. merchant-center [arg]
{{ROTATION_TITLE}}	display name, e.g. Merchant Center [interview]
{{PERSONA}} / {{PERSONA_VIBE}}	soul name and one line [interview]
{{RUNBOOK_INDEXES}}	index pages table [interview + descendants]
{{CLOUD_ID}}	Confluence cloudId [getAccessibleAtlassianResources]
{{ENV_TABLE}}	env to cloud id to log account [grafana_list_datasources]
{{SERVICE_VOCAB}} / {{SIGNAL_VOCAB}}	matching vocab [ingest]
{{SERVICE_REPOS}}	service to repo map [runbook metadata + interview]
{{DASHBOARDS}}	standing dashboard UIDs [interview + Grafana search]
{{LOGSCALE_PACKAGE}}	humio-packages path [interview]
{{ESCALATION_DEFAULT}}	default escalation path [incident.io]
{{ALERT_PHRASINGS}}	trigger phrases for the description [interview]
{{SCOPE_TAGS}}	scope, usually monitoring and a team tag [interview]

Guardrails

• This skill writes files. It never runs operational commands. Read only gh and MCP reads only.
• The generated skill inherits the safety rules. Read only. Suggests operational commands, never runs them. Prompt free allowlist documented in its README.
• Do not ship a generated skill without the human review and the incident spot check in step 8.

Examples

# Scaffold an on-call skill for the Frontend Platform rotation
/on-call-skill-creator frontend-platform

# Adopt it for Checkout
/on-call-skill-creator checkout