Agent Skills

A collection of installable Agent Skills for AI coding CLIs (Codex, Claude Code, QoderWork, Qoder CLI). Each skill is a self-contained directory with a SKILL.md manifest and optional supporting files that are loaded on demand.

Skills

Skill	Description	Install path examples
loop-verify	Lightweight outcome/rubric verification-loop coding workflow with E2E scope discovery, long-task progress artifacts, root-cause, persistence, user-flow evidence, and optional external review gates	Codex project skill: .agents/skills/loop-verify Claude user skill: ~/.claude/skills/loop-verify
golang-ddd	Chinese-language Go DDD design, implementation, refactoring, and review guidance covering strategic/tactical modeling, clean boundaries, repositories, transactions, outbox, idempotency, CQRS, testing, and pragmatic complexity control	Codex/Agent user skill: ~/.agents/skills/golang-ddd Project skill: .agents/skills/golang-ddd
md2docx	Markdown to Word (DOCX) converter with tables, images, CJK fonts, Mermaid	~/.claude/skills/md2docx or project skill directory
confluence-publish	Publish local HTML files to Confluence 6.x wiki pages via browser automation (agent-browser + Chrome CDP), with HTML simplification, Unicode sanitization, TinyMCE injection, and optional one-click REST API publish	Codex/Agent user skill: ~/.agents/skills/confluence-publish
zstack-gsstack-ops	Sanitized GSStack/ZStack lab operations workflow for snapshot recovery, KubeKey replayability, and safe E2E verification	Codex user skill: ~/.agents/skills/zstack-gsstack-ops
ssh-reuse	Reusable OpenSSH operations with ControlMaster/ControlPersist, ssh-agent key auth, ProxyJump, tmux sessions, resumable transfer, batch commands, server-to-server copy, and tunnels	Codex user skill: ~/.agents/skills/ssh-reuse

Machine-local skills that reference private lab infrastructure (gsstack-local-dev, gh-ceec-local-dev) live in a separate private repository and read lab-specific values from local ignored env files. | container-e2e-browser | bb-browser automation patterns for GSStack container module E2E testing: login flow, navigation, form interaction, kubectl verification, error recovery, and a complete end-to-end walkthrough | Codex user skill: ~/.agents/skills/container-e2e-browser | | container-e2e-full | Step-by-step full E2E test flows for all GSStack container modules (workloads, network, storage, RBAC, Ingress, NetworkPolicy, HPA, XSKY CSI, components, observability) with YAML templates, troubleshooting, and functional verification | Codex user skill: ~/.agents/skills/container-e2e-full |

Quick Install

Clone this repository first:

git clone https://github.com/RoshanDev/skills.git
cd skills

Preferred symlink install

Keep this repository as the canonical source and link user/project skill paths to it. This avoids silent drift between root-level, user-level, and project-level copies.

Replace /path/to/skills with your local clone path, for example $HOME/Developer/skills.

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/loop-verify ~/.agents/skills/loop-verify
ln -sfn /path/to/skills/loop-verify ~/.codex/skills/loop-verify
ln -sfn /path/to/skills/golang-ddd ~/.agents/skills/golang-ddd
ln -sfn /path/to/skills/golang-ddd ~/.codex/skills/golang-ddd
ln -sfn /path/to/skills/zstack-gsstack-ops ~/.agents/skills/zstack-gsstack-ops
ln -sfn /path/to/skills/zstack-gsstack-ops ~/.codex/skills/zstack-gsstack-ops
ln -sfn /path/to/skills/ssh-reuse ~/.agents/skills/ssh-reuse
ln -sfn /path/to/skills/ssh-reuse ~/.codex/skills/ssh-reuse
ln -sfn /path/to/skills/container-e2e-browser ~/.agents/skills/container-e2e-browser
ln -sfn /path/to/skills/container-e2e-browser ~/.codex/skills/container-e2e-browser
ln -sfn /path/to/skills/container-e2e-full ~/.agents/skills/container-e2e-full
ln -sfn /path/to/skills/container-e2e-full ~/.codex/skills/container-e2e-full
ln -sfn /path/to/skills/confluence-publish ~/.agents/skills/confluence-publish
ln -sfn /path/to/skills/confluence-publish ~/.codex/skills/confluence-publish

For a project-local skill path, prefer the same symlink unless the project has a genuinely different contract. If a project copy is modified, merge the reusable part back here before relying on it.

loop-verify

Install the whole skill directory, not only SKILL.md. The skill links to examples.md, reference.md, e2e-scope-discovery.md, user-flow-evidence.md, external-review.md, long-task-progress.md, and outcomes.md, so copying only the manifest will break progressive disclosure.

Codex project skill

From your target project root:

mkdir -p .agents/skills
ln -sfn /path/to/skills/loop-verify .agents/skills/loop-verify

Then invoke it in Codex with:

$loop-verify contract
$loop-verify outcome
$loop-verify e2e-scope
$loop-verify user-flow
$loop-verify execute
$loop-verify review

Claude Code user skill

mkdir -p ~/.claude/skills
ln -sfn /path/to/skills/loop-verify ~/.claude/skills/loop-verify

QoderWork

mkdir -p ~/.qoderwork/skills
ln -sfn /path/to/skills/loop-verify ~/.qoderwork/skills/loop-verify

golang-ddd

Install the whole directory because the main manifest progressively loads the reference guides, templates, validation scripts, and compilable order example.

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/golang-ddd ~/.agents/skills/golang-ddd
ln -sfn /path/to/skills/golang-ddd ~/.codex/skills/golang-ddd

For a project-local installation:

mkdir -p .agents/skills
ln -sfn /path/to/skills/golang-ddd .agents/skills/golang-ddd

Validate the skill and its example from the repository root:

python3 golang-ddd/scripts/validate_skill.py golang-ddd
python3 golang-ddd/scripts/check_domain_imports.py golang-ddd/examples/order
(cd golang-ddd/examples/order && go test -race ./...)

zstack-gsstack-ops

The public skill is intentionally sanitized. Keep lab endpoints, private IDs, credentials, and private environment details in ignored local files or project-private documentation by default. For internal lab execution, those private local sources are valid inputs. If the user explicitly selects trusted-lab-unredacted, private notes, private repo commits, screenshots, HAR request bodies, command output, and final chat answers may keep raw values while staying inside that trusted context.

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/zstack-gsstack-ops ~/.agents/skills/zstack-gsstack-ops
ln -sfn /path/to/skills/zstack-gsstack-ops ~/.codex/skills/zstack-gsstack-ops

ssh-reuse

Use this skill for repeated SSH debugging through OpenSSH host aliases, ssh-agent, ControlMaster, ControlPersist, ProxyJump, tmux sessions, resumable transfer, batch commands, server-to-server copy, and local tunnels. Keep real server details in ~/.ssh/config or ~/.ssh/servers.d/*.conf, not in this public repository.

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/ssh-reuse ~/.agents/skills/ssh-reuse
ln -sfn /path/to/skills/ssh-reuse ~/.codex/skills/ssh-reuse

container-e2e-browser / container-e2e-full

Both skills are sanitized: image registries use harbor.example.com placeholders. Keep the real registry, frontend URL, admin credentials, cookie/session files, browser profiles, storage state, and kubeconfig in ignored local env files or project-private docs by default. In internal labs, agents should use those local private sources and real browser sessions directly instead of inventing API token injection paths. With explicit trusted-lab-unredacted approval, private evidence and final chat answers may include the raw values.

Install the whole skill directory, not only SKILL.md. Each skill links to supporting files for progressive disclosure:

• container-e2e-browser references walkthrough.md (complete end-to-end example) and troubleshooting.md (error recovery)
• container-e2e-full references yaml-templates.md (all test resource YAMLs), http-workload-service.md (known-good HTTP workload + Service for curl verification), troubleshooting.md (failure diagnosis), and xsky-module.md (XSKY CSI lifecycle testing)

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/container-e2e-browser ~/.agents/skills/container-e2e-browser
ln -sfn /path/to/skills/container-e2e-browser ~/.codex/skills/container-e2e-browser
ln -sfn /path/to/skills/container-e2e-full ~/.agents/skills/container-e2e-full
ln -sfn /path/to/skills/container-e2e-full ~/.codex/skills/container-e2e-full

md2docx

cd md2docx
./install.sh

confluence-publish

Install the whole skill directory. The scripts need powershell.exe (WSL), Windows Chrome, agent-browser on Windows PATH, and beautifulsoup4.

mkdir -p ~/.agents/skills ~/.codex/skills
ln -sfn /path/to/skills/confluence-publish ~/.agents/skills/confluence-publish
ln -sfn /path/to/skills/confluence-publish ~/.codex/skills/confluence-publish

Directory Structure

skills/
├── golang-ddd/
│   ├── SKILL.md
│   ├── README.md
│   ├── assets/
│   ├── examples/order/
│   ├── references/
│   └── scripts/
├── loop-verify/
│   ├── SKILL.md
│   ├── e2e-scope-discovery.md
│   ├── examples.md
│   ├── external-review.md
│   ├── long-task-progress.md
│   ├── outcomes.md
│   ├── reference.md
│   └── user-flow-evidence.md
├── zstack-gsstack-ops/
│   ├── SKILL.md
│   ├── agents/
│   │   └── openai.yaml
│   └── scripts/
│       └── zstack_create_initial_snapshots.py
├── ssh-reuse/
│   ├── SKILL.md
│   ├── config-guide.md
│   ├── commands.md
│   ├── troubleshooting.md
│   └── scripts/
│       └── sshx
├── container-e2e-browser/
│   ├── SKILL.md
│   ├── walkthrough.md
│   └── troubleshooting.md
├── container-e2e-full/
│   ├── SKILL.md
│   ├── http-workload-service.md
│   ├── yaml-templates.md
│   ├── troubleshooting.md
│   └── xsky-module.md
├── md2docx/
│   ├── SKILL.md
│   ├── converter-template.md
│   ├── install.sh / install.ps1
│   └── examples/
└── confluence-publish/
    ├── SKILL.md
    ├── troubleshooting.md
    └── scripts/
        ├── publish_confluence_html.py
        └── check_confluence_meta.py

Design Notes

• Keep the main SKILL.md under control and push detailed examples/checklists/rubric patterns into supporting files.
• Do not duplicate the same workflow across multiple tools. Prefer one owner for intent, one owner for evidence, and explicit gates for verification.
• For coding workflows, use acceptance criteria as the contract and tests/checks as executable evidence. Do not let the implementing agent self-certify without command output, AC/rubric coverage, E2E scope evidence when relevant, user-flow evidence when relevant, and persistence status when relevant.
• For long-running work, persist task state outside chat context using progress artifacts such as progress.md and feature_list.json.
• For feature work, the agent should infer the E2E Impact Map from code/routes/UI/API/tests and ask only for genuinely missing business or environment decisions.
• For UI/browser defects, curl/hooks/Python/direct API checks are diagnosis/setup evidence only. Final PASS needs browser-driven user-path evidence unless explicitly waived.
• External review tools such as language-specific review skills or open-code-review are optional advisory/fresh-review inputs. They do not replace loop-verify gates.

License

MIT - see LICENSE.