Operational Skills
A 163-skill operational corpus is tracked at
BagelHole/DevOps-Security-Agent-Skills
(MIT, © Toby Miller). Agent Compass keeps no copy of it. It installs a curated
146 of them on request, and rewrites each one on the way in.
That rewrite is the reason this skill exists. Upstream examples put secrets in
argv, and a process list is readable by every other local process. The
installer replaces eight such passages and prepends a safety gate to every
skill. Read an operational skill only after installing it through Agent Compass —
the raw upstream text is not the reviewed text.
Procedure
1. Install the skill the task needs
# One skill or a few
agent-compass external-skills . --source devops-security --skill kubernetes-ops,helm-charts
# The whole curated set, for a platform or security engagement
agent-compass external-skills . --source devops-security --recommended
# User-wide instead of per-project
agent-compass external-skills --source devops-security --recommended --global
This writes .claude/skills/ for Claude Code and .agents/skills/ for Codex and
Copilot, plus a Copilot instructions file, and drops the MIT notice beside them.
Confirm the output says it applied the Agent Compass safety gate and
argv-secret narrowings. If that line is missing, the installed text is not the
reviewed text — stop and report it.
2. Pick by area, not by name
agent-compass skills-sync --list-packs prints every capability pack. Rough map:
| Task area |
Pack |
| CI/CD, containers, Kubernetes, observability, releases |
devops-platform |
| Scanning, secrets, hardening, network security, incident response |
security |
| Cloud provisioning, Terraform, networking, storage, databases |
infrastructure |
| SOC 2, ISO 27001, GDPR, HIPAA, PCI DSS, FedRAMP, audit |
compliance |
| Model serving, GPU, RAG, inference scaling, LLM gateways |
the AI-operations subpacks |
Install one pack's skills, not the corpus. A single cloud or a single compliance
framework is usually the right scope — the focused subpacks exist for that.
The 17 skills Agent Compass does not curate are still tracked and installable by
name. They are uncurated: read one before you act on it.
3. The gate is not advisory
Every installed skill carries this, and it outranks the skill's own examples:
- Authorization first. Confirm the exact target: environment, account,
cluster, namespace, repository, and data classification. A skill that says
"run this against prod" does not grant that permission.
- Read-only first. Use plan, diff, check, or dry-run modes before any
mutation. Never deploy, delete, rotate a credential, fail over, contain, or
write to production without explicit approval in the conversation.
- Preserve rollback and evidence. Back up state before a destructive step.
During an incident, collect evidence before remediation when it is safe to.
- Least privilege, no secret in the clear. Never print, commit, or paste a
secret into a prompt, a log, a command, or an example. If an upstream example
still shows a secret in
argv after install, that is a bug in the narrowings —
report it rather than running it.
- Verify against current docs. Flags, API versions, and control names age.
The vendor's current documentation outranks the skill.
- Compliance skills prepare, they do not certify. No output of this corpus is
certification, attestation, or legal advice.
Full detail: operational-safety.
4. Validate what you changed
An operational change is validated by the tool that owns the state, not by the
skill's prose. Run the real command and report its output:
terraform plan # never report `apply` you did not run
kubectl diff -f <manifest>
helm upgrade --dry-run
Report passed / failed / partial / not run per command and say whether a
failure is pre-existing. AGENTS.md §3 and §4.
5. Plan before an operational change
Anything that touches shared or production state gets a plan first: the target,
the blast radius, the rollback, and the approval. See
plan-before-operational-change.
Freshness
agent-compass upstream-skills --check-updates
agent-compass upstream-skills --update devops-security --dry
An install is a snapshot of a pin, so it is recorded in
.agent/external-skills.json. When the compass pin moves, the corrected text you
installed is out of date — and here that includes the safety gate and the
narrowings. You do not have to remember: the session-start check, recommend,
and install --doctor all report it.
agent-compass external-skills . --check # offline; which installs are behind
agent-compass external-skills . --upgrade # re-install them at the current pin
Related
- operational-skills.md — the eight
narrowings in full, the gate text, and the tracked inventory.
- operational-safety — the rules
the gate summarises.
runbook-creation, incident-management, and disaster-recovery are part of
the tracked corpus, not compass skills. Install one before you rely on it.
1---2name: operational-skills3description: Use when the task is DevOps, cloud, Kubernetes, containers, CI/CD, observability, security scanning, secrets, hardening, networking, storage, databases, incident response, LLM/GPU serving, or a compliance framework (SOC 2, ISO 27001, GDPR, HIPAA, PCI DSS, FedRAMP). Installs the pinned operational skill for the task with the Agent Compass safety gate applied, then holds the authorization and production-write rules in force. Triggers: terraform, kubernetes, kubectl, helm, argocd, docker, github actions, gitlab ci, prometheus, grafana, opentelemetry, vault, secrets manager, hardening, vulnerability scan, SBOM, incident, runbook, disaster recovery, vllm, ollama, soc2, iso27001, gdpr, hipaa, pci.4---56# Operational Skills78A 163-skill operational corpus is tracked at9[`BagelHole/DevOps-Security-Agent-Skills`](https://github.com/BagelHole/DevOps-Security-Agent-Skills)10(MIT, © Toby Miller). Agent Compass keeps no copy of it. It installs a curated11146 of them on request, and **rewrites each one on the way in**.1213That rewrite is the reason this skill exists. Upstream examples put secrets in14`argv`, and a process list is readable by every other local process. The15installer replaces eight such passages and prepends a safety gate to every16skill. Read an operational skill only after installing it through Agent Compass —17the raw upstream text is not the reviewed text.1819## Procedure2021### 1. Install the skill the task needs2223```bash24# One skill or a few25agent-compass external-skills . --source devops-security --skill kubernetes-ops,helm-charts2627# The whole curated set, for a platform or security engagement28agent-compass external-skills . --source devops-security --recommended2930# User-wide instead of per-project31agent-compass external-skills --source devops-security --recommended --global32```3334This writes `.claude/skills/` for Claude Code and `.agents/skills/` for Codex and35Copilot, plus a Copilot instructions file, and drops the MIT notice beside them.3637Confirm the output says it **applied the Agent Compass safety gate and38argv-secret narrowings**. If that line is missing, the installed text is not the39reviewed text — stop and report it.4041### 2. Pick by area, not by name4243`agent-compass skills-sync --list-packs` prints every capability pack. Rough map:4445| Task area | Pack |46| --------- | ---- |47| CI/CD, containers, Kubernetes, observability, releases | `devops-platform` |48| Scanning, secrets, hardening, network security, incident response | `security` |49| Cloud provisioning, Terraform, networking, storage, databases | `infrastructure` |50| SOC 2, ISO 27001, GDPR, HIPAA, PCI DSS, FedRAMP, audit | `compliance` |51| Model serving, GPU, RAG, inference scaling, LLM gateways | the AI-operations subpacks |5253Install one pack's skills, not the corpus. A single cloud or a single compliance54framework is usually the right scope — the focused subpacks exist for that.5556The 17 skills Agent Compass does not curate are still tracked and installable by57name. They are uncurated: read one before you act on it.5859### 3. The gate is not advisory6061Every installed skill carries this, and it outranks the skill's own examples:6263- **Authorization first.** Confirm the exact target: environment, account,64 cluster, namespace, repository, and data classification. A skill that says65 "run this against prod" does not grant that permission.66- **Read-only first.** Use plan, diff, check, or dry-run modes before any67 mutation. Never deploy, delete, rotate a credential, fail over, contain, or68 write to production without explicit approval in the conversation.69- **Preserve rollback and evidence.** Back up state before a destructive step.70 During an incident, collect evidence before remediation when it is safe to.71- **Least privilege, no secret in the clear.** Never print, commit, or paste a72 secret into a prompt, a log, a command, or an example. If an upstream example73 still shows a secret in `argv` after install, that is a bug in the narrowings —74 report it rather than running it.75- **Verify against current docs.** Flags, API versions, and control names age.76 The vendor's current documentation outranks the skill.77- **Compliance skills prepare, they do not certify.** No output of this corpus is78 certification, attestation, or legal advice.7980Full detail: [operational-safety](../../docs/guidelines/operational-safety.md).8182### 4. Validate what you changed8384An operational change is validated by the tool that owns the state, not by the85skill's prose. Run the real command and report its output:8687```bash88terraform plan # never report `apply` you did not run89kubectl diff -f <manifest>90helm upgrade --dry-run91```9293Report `passed` / `failed` / `partial` / `not run` per command and say whether a94failure is pre-existing. `AGENTS.md` §3 and §4.9596### 5. Plan before an operational change9798Anything that touches shared or production state gets a plan first: the target,99the blast radius, the rollback, and the approval. See100[`plan-before-operational-change`](../../knowledge/instincts/plan-before-operational-change.md).101102## Freshness103104```bash105agent-compass upstream-skills --check-updates106agent-compass upstream-skills --update devops-security --dry107```108109An install is a snapshot of a pin, so it is recorded in110`.agent/external-skills.json`. When the compass pin moves, the corrected text you111installed is out of date — and here that includes the safety gate and the112narrowings. You do not have to remember: the session-start check, `recommend`,113and `install --doctor` all report it.114115```bash116agent-compass external-skills . --check # offline; which installs are behind117agent-compass external-skills . --upgrade # re-install them at the current pin118```119120## Related121122- [operational-skills.md](../../docs/tooling/operational-skills.md) — the eight123 narrowings in full, the gate text, and the tracked inventory.124- [operational-safety](../../docs/guidelines/operational-safety.md) — the rules125 the gate summarises.126- `runbook-creation`, `incident-management`, and `disaster-recovery` are part of127 the tracked corpus, not compass skills. Install one before you rely on it.
me-cedric