Skip to content
Back to skills

Ansible Designer

ASecurity

AI-assisted Ansible authoring toolkit for Claude Code. Scaffolds, reviews, and updates playbooks, roles, collections, and ansible.cfg files following production best practices. Sub-commands: new-playbook, review-playbook, update-playbook, new-role, review-role, update-role, new-collection, review-collection, update-collection, new-conf, review-conf, update-conf. Runs discovery (CLAUDE.md to ansible.cfg to README to filesystem) at the start of every command.

  • 5 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added May 26, 2026
devopsgobashtestingapisecurity

Works with

  • claude code
  • api

Security analysis

A100/100

Pro scans all 11 files and shows the line behind each finding

Scanned May 27, 2026

npx -y skills add 3A2DEV/ansible-designer --skill ansible-designer --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ansible Designer?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ansible Designer
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/3a2dev-ansible-designer/badge)](https://www.skillsdirectory.com/skills/3a2dev-ansible-designer)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ansible-designer
description: "AI-assisted Ansible authoring toolkit for Claude Code. Scaffolds, reviews, and updates playbooks, roles, collections, and ansible.cfg files following production best practices. Sub-commands: new-playbook, review-playbook, update-playbook, new-role, review-role, update-role, new-collection, review-collection, update-collection, new-conf, review-conf, update-conf. Runs discovery (CLAUDE.md to ansible.cfg to README to filesystem) at the start of every command."
---

# ansible-designer

AI-assisted Ansible authoring toolkit. Scaffolds, reviews, and updates Ansible projects with production-quality output following official ansible-core 2.15+ conventions.

---

## Available Commands

All commands are top-level regardless of install path.

### Playbook Commands
| Command | Description |
|---------|-------------|
| `/new-playbook` | Create a new playbook (site, component, or AWX-ready) |
| `/review-playbook` | Review an existing playbook — severity report, no file modification |
| `/update-playbook` | Update a playbook — shows diff, requires confirmation |

### Role Commands
| Command | Description |
|---------|-------------|
| `/new-role` | Scaffold a complete role — asks about multi-OS support |
| `/review-role` | Review a role — severity report, no file modification |
| `/update-role` | Update a role — shows diff, requires confirmation |

### Collection Commands
| Command | Description |
|---------|-------------|
| `/new-collection` | Scaffold a new collection with galaxy.yml, plugins, roles structure |
| `/review-collection` | Review a collection — severity report, no file modification |
| `/update-collection` | Update a collection — shows diff, requires confirmation |

### ansible.cfg Commands
| Command | Description |
|---------|-------------|
| `/new-conf` | Generate an annotated ansible.cfg for dev, CI, or AWX |
| `/review-conf` | Review an ansible.cfg — severity report, no file modification |
| `/update-conf` | Update ansible.cfg — shows diff, requires confirmation |

---

## Global Rules

Every sub-command enforces these rules without exception:

1. **Discovery first** — At command start, read in order: `CLAUDE.md` → `ansible.cfg` → `README.md` → filesystem scan. Build internal context (roles, collections, paths, namespace). Skip if user already provided all required parameters inline. See `references/discovery.md`.

2. **Never overwrite silently** — Before writing any file, show a summary (new files) or unified diff (modifications). Wait for explicit user confirmation (`yes` / `y`). Only write after confirmation.

3. **FQCN mandatory** — Every module reference uses the Fully Qualified Collection Name. `ansible.builtin.copy`, never `copy`. `ansible.builtin.service`, never `service`. See `references/best_practices.md`.

4. **no_log: true on secrets** — Every task handling passwords, tokens, API keys, vault variables, or credentials must include `no_log: true`. See `references/security_vault.md`.

5. **Tags on every task** — Minimum: component name + action category (`install`, `configure`, `service`, `validate`, `security`, `cleanup`). No task may be untagged.

6. **review never modifies** — `review-*` commands produce a structured severity report (CRITICAL / WARNING / INFO) only. They never write, modify, or suggest `sed` commands. Report only.

7. **update always diffs** — `update-*` commands read the existing file, compute the change, show a unified diff, and wait for explicit confirmation before writing a single byte.

8. **Show file tree after writes** — After any write operation, list all created or modified files and display the resulting file tree.

9. **Testing-aware output** — New or updated examples must include a realistic validation path. See `references/testing.md`.

10. **Suggest next step** — End every command with a concrete suggestion: which command to run next, or what to validate.

---

## Standard Operational Flow

Every command follows this exact sequence:

### Step 1 — Discovery
```
Read: CLAUDE.md → ansible.cfg → README.md → filesystem scan
Build: internal context (roles_path, collections_path, existing roles, collections, inventory)
Report: "Discovery complete: [summary of what was found]"
```

Skip Step 1 only if the user provided all required parameters inline.

### Step 2 — Parameter Collection
- If the user already provided all required parameters: proceed to Step 3.
- Otherwise: ask **one question at a time**, using discovery context for smart defaults.
- Never ask for something that can be inferred from discovery (e.g., don't ask for namespace if a collection already exists).

### Step 3 — Pre-Write Confirmation
- For **new** files: show a summary of what will be created (paths + brief description).
- For **update** commands: show a unified diff (`--- original`, `+++ proposed`).
- **Wait for explicit user confirmation** (`yes`, `y`, or equivalent) before proceeding.
- If user says no: ask what to change and loop back to Step 2.

### Step 4 — Execution
- Write or modify files using bash commands.
- Follow all global rules: FQCN, tags, no_log.
- Use templates from the appropriate `references/` file as the base.

### Step 5 — Final Output
```
Show file tree of all created/modified files.
Suggest: "Next step: [specific actionable suggestion]"
```

---

## Reference Files

| File | Used by |
|------|---------|
| `references/discovery.md` | All commands — Step 1 |
| `references/best_practices.md` | All commands — FQCN, tags, no_log, idempotency |
| `references/playbook.md` | new-playbook, review-playbook, update-playbook |
| `references/role.md` | new-role, review-role, update-role |
| `references/collection.md` | new-collection, review-collection, update-collection |
| `references/ansible_cfg.md` | new-conf, review-conf, update-conf |
| `references/inventory.md` | All commands — inventory context |
| `references/testing.md` | All commands — validation and example test guidance |
| `references/security_vault.md` | All commands — secret handling and vault-safe patterns |
| `references/plugins.md` | Collection commands — module/filter/lookup guidance |

---

## Installation

Install via the Claude Code marketplace from `3A2DEV/ansible-designer`. This preserves the `ansible-designer:` namespace prefix for all sub-commands.

Requires Claude Code.

Files in this skill

  • SKILL.md6.3 KB
  • references/ansible_cfg.md3.1 KB
  • references/best_practices.md6.2 KB
  • references/collection.md2.8 KB
  • references/discovery.md4.3 KB
  • references/inventory.md2.1 KB
  • references/playbook.md4.6 KB
  • references/plugins.md1.4 KB
  • references/role.md4.1 KB
  • references/security_vault.md1.5 KB
  • references/testing.md1.4 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.