oh-my-cloud-skills
Ten plugins for Claude Code and
Codex: cloud content, AWS operations,
agent and plugin conversion, peer review, implementation delegation, project documentation,
concise response delivery, and PR and agent status views.
Documentation ·
Changelog ·
Releases ·
License
Plugins
| Plugin | Workflows |
|---|
| aws-content-plugin | Interactive web decks, editable PowerPoint, architecture and animated diagrams, technical documents, GitBook sites, workshops, brochures, and personal profiles |
| aws-ops-plugin | AWS/EKS troubleshooting, health checks, networking, identity, observability, storage, databases, analytics, cost, and Well-Architected reviews |
| kiro-power-converter | Convert plugin sources or individual skills into Kiro Powers, including steering, hooks, assets, and MCP configuration |
| agentcore-creator | Discover requirements, design and test locally, then prepare AgentCore harness configuration or a Runtime application |
| co-agent | Peer review, decisions, ADRs, context synchronization, consensus/harness implementation, PR feedback fixes, and ADR reconciliation |
| project-init | Initialize and maintain the current host's project instructions, skills, architecture docs, ADRs, runbooks, and reference guides |
| kiro | Delegate implementation to Kiro CLI, with host verification and optional commit/push review and web search |
| token-saver | Concise response guidance with full reasoning, verification, complete artifacts, and required report formats |
| pr-dash | A Claude Code pane and status line for a pull request's CI checks and AI review, plus a read-only status skill for Codex |
| agent-dash | A Claude Code pane and status line for the session's agents, who delegated to whom and who is waiting, with an opt-in subagent model policy, plus a read-only roster skill for Codex |
Both Claude and
Codex marketplaces contain these plugins.
Manifests and generated inventories define their current entry points; source
skills, commands, agents, and generated Codex entries are different populations.
Shared procedures have one maintained source. Host settings are independent under
hosts/claude/ and hosts/codex/; each marketplace selects its own self-contained
packages/<host>/ output. See host packaging.
Installation
Claude Code
Run these commands inside Claude Code, installing only the plugins you need:
/plugin marketplace add Atom-oh/oh-my-cloud-skills
/plugin install aws-content-plugin@oh-my-cloud-skills
/plugin install co-agent@oh-my-cloud-skills
Use another name from the plugin table to install that package. From a local checkout:
python3 scripts/sync-codex-plugins.py
claude --plugin-dir ./packages/claude/aws-content-plugin
Remove a package with /plugin uninstall aws-content-plugin@oh-my-cloud-skills;
remove the marketplace with /plugin marketplace remove oh-my-cloud-skills.
Codex CLI
Use a Codex CLI with plugin support:
codex plugin marketplace add Atom-oh/oh-my-cloud-skills
codex
Open /plugins, select Oh My Cloud Skills, and install the desired packages.
For local development, regenerate packages with python3 scripts/sync-codex-plugins.py,
then run codex plugin marketplace add ./ from this repository.
Manage the registered source with:
codex plugin marketplace list
codex plugin marketplace upgrade oh-my-cloud-skills
codex plugin marketplace remove oh-my-cloud-skills
Uninstall individual packages through /plugins. Start a new thread when needed
to load the installed entries.
Skills, commands, and specialists in Codex
Request a workflow naturally or select it with /skills or the $ picker.
Each plugin's .codex-plugin/inventory.json maps generated entries to shared
procedures. Same-name aliases can share an entry. Project-init health-check uses
source-command-health-check.
Claude commands become skill workflows in Codex. Agent Markdown supplies specialist
instructions; it does not register native Codex agent types, models, memory, or
permission grants. Internal specialist entries require explicit host selection.
External CLI workflows still need the relevant executable and authentication.
Declared plugin hooks require host trust. Project-init instead provides project
hook templates: install them in the consumer project, establish project trust,
inspect their definitions, and grant hook trust. Installing project-init alone
does not activate those project hooks. See
runtime verification and evidence scope.
Quick Start
| Request | Workflow |
|---|
| “Create a 30-minute English EKS operations deck with speaker notes and a quiz.” | reactive-presentation |
| “Create an editable AWS light PowerPoint deck from this outline.” | aws-light-fcd |
| “Draw this multi-AZ AWS architecture and export PNG.” | architecture-diagram |
| “Show request traffic and a failover scenario.” | animated-diagram |
| “Build a Workshop Studio lab with verification and cleanup.” | workshop-creator |
| “Diagnose why these pods cannot reach the service.” | ops-network-diagnosis |
| “Assess this cluster's health.” | ops-health-check |
| “Get a second opinion on this diff.” | co-agent (review mode) |
| “Delegate this approved implementation plan to Kiro.” | kiro-delegate |
| “Initialize this existing project for Codex.” | init-project |
| “Use concise replies while completing the full task.” | concise-responses |
Commands shown below name the Claude workflows; select their corresponding
installed skill entries in Codex. Review a concrete deployment or publication
candidate before executing actions that require authorization.
Reactive Presentation
The reactive-presentation skill
authors Remarp source and builds interactive HTML. It supports fragments, presenter
notes, comparison/tabs, quizzes, timelines, checklists, code, and custom calculators
or simulations. A multi-file deck uses _presentation.md, block files, assets,
and generated HTML with a shared framework.
A local source workflow, run from this marketplace checkout. For an English directory
build, set lang: en in my-presentation/_presentation.md frontmatter:
python3 plugins/aws-content-plugin/skills/reactive-presentation/scripts/remarp_to_slides.py validate ./my-presentation/
python3 plugins/aws-content-plugin/skills/reactive-presentation/scripts/remarp_to_slides.py build ./my-presentation/
python3 plugins/aws-content-plugin/skills/reactive-presentation/scripts/remarp_to_slides.py sync ./my-presentation/
Use Remarp quick start,
CLI reference, and
keyboard controls for the
current syntax and runtime behavior. The
VS Code extension supplies preview,
source navigation, build, and issue annotations; slide-fix applies those annotations.
PPTX theme extraction applies a supplied brand to web slides. Screenshot-based
PPTX export captures a built deck; native editable PowerPoint uses
aws-light-fcd.
Keep these output formats distinct. Shared AWS icons are reused across the workflows.
Architecture Diagrams
The architecture-diagram skill
provides YAML-driven Draw.io layout, hand-authored XML for unsupported structures,
and Excalidraw sketch output. Validate XML and require layout score at least 80
before Draw.io export. Sizes, colors, labels, and spacing come from the
canonical design tokens.
Use simple Canvas DSL for a small linear flow; use HTML/CSS for larger or grouped
slide diagrams. Remarp's :::archify path renders through its configured, pinned
Archify dependency and embeds an explorable diagram. Check the source skill for
that dependency and output contract rather than treating it as another marketplace plugin.
Animated Diagrams
animated-diagram
uses SVG/SMIL for repeated traffic motion and JavaScript/CSS state machines for
controlled scaling, deployment, and failover scenarios. Keep a readable static
architecture, labels, legend, and working reset/replay controls.
Documents
The document agent writes
technical reports and solution comparisons with evidence, diagrams, and useful
references. brochure creates a
product or solution landing page; gh-home
creates a personal profile or portfolio. Both produce responsive HTML.
GitBook Sites
gitbook supplies site structure,
SUMMARY navigation, rich components, diagrams, and cross-link checks. Create language
variants only when the deliverable requires them.
Workshops
workshop-creator
produces Workshop Studio modules, labs, directives, and optional infrastructure.
Each lab explains prerequisites, commands, expected results, verification, and cleanup.
AWS Ops
Use ops-troubleshoot for a
concrete failure, ops-health-check
for an overall assessment, and the network, observability, security-audit, or
Well-Architected skill for its specific scope. Specialist procedures cover EKS,
network, IAM, observability, storage, database, analytics, cost, and Well-Architected reviews;
the coordinator correlates multi-domain incidents.
The workflow is scope → read-only evidence → diagnosis → authorized remediation →
verification. A scored assessment must identify its rubric and evidence gaps.
Bundled MCP servers are defined in the host manifests; additional integrations
require their own setup. See the operations guide.
Kiro Power Converter
kiro-convert accepts a
local plugin, GitHub source, marketplace candidate, or individual skill. It maps
metadata, steering, hooks, MCP settings, and assets into Kiro Power format.
python3 plugins/kiro-power-converter/skills/kiro-convert/scripts/convert_plugin_to_power.py \
--source ./plugins/aws-ops-plugin --output /var/tmp/aws-ops-power --target export
Targets include export, project, and global installation. --preserve-skills
retains supported skill structure. Resolve ambiguous marketplace candidates to an
explicit source; inspect generated metadata and required environment variables
before loading the Power in Kiro.
AgentCore Creator
agentcore-create follows
Discovery → Design → Skill-First Build → Convert → Deploy and Verify. Existing
plugin input can enter at conversion. Choose harness configuration for a supported
managed loop or generated Strands/Runtime code for custom orchestration. Model
mapping, compatibility, tool/Gateway integration, and Memory options live in the
converter and its references. Generated files are not proof of cloud deployment.
Co-agent
The current host chairs, excludes itself from external peer selection, verifies
findings against source, and owns the final synthesis. Peer responses are advisory;
agreement alone does not validate a finding.
The co-agent skill has six modes:
| Mode | Result |
|---|
review | Evidence-checked findings and disagreements |
decide | Options, tradeoffs, and a host recommendation |
adr | A decision record informed by peer perspectives |
sync-context | Marked AGENTS.md distilled from CLAUDE.md, plus the Kiro steering bridge |
consensus | Host implementation with required peer gates |
harness | Eligible peer implementation in isolated worktrees; host design, verification, and commits |
/co-agent:setup is a separate readiness command. It probes actual peer
usability; an installed binary alone is insufficient. Casual review/decide/ADR can
continue solo with an explicit notice. Consensus and harness require the usable
peer evidence and coverage their gates specify, and cannot silently become solo.
A plain code-review request does not automatically authorize multi-AI fan-out.
/co-agent:setup
/co-agent review the current diff
/co-agent:configure
/co-agent:sync-context
/co-agent:configure shows effective models, supported effort, profiles, timeouts,
and budgets. The canonical defaults
and local overrides are authoritative; provider catalog IDs need not match across vendors.
pr-autofix polls AI/human feedback,
plans confirmed fixes, applies them in an isolated worktree, validates, and pushes
within configured bounds. decision-reconcile
checks ADR contradictions and drift, then drafts a superseding decision.
Kiro Delegation
/kiro:setup probes the CLI and prepares its agents. /kiro:delegate follows a
host-authored plan through task worktrees, Kiro implementation, diff capture, scope
validation, and host tests/commits. /kiro:review requests a review;
/kiro:configure shows effective settings.
The scope guard controls which captured changes reach the main tree. It does not
sandbox the Kiro process. Enabling the implementer's shell tool is a separate trust
decision. Default delegation, automatic review, and delegated search have separate
opt-in settings. Review settings are independent of implementation settings; see
kiro.defaults.json.
Project Init
Project-init adapts the existing repository to the selected host. Its workflows
include init-project, sync-docs, add-adr, add-module, add-runbook,
add-reference-doc, generate-readme, generate-changelog, and health-check.
Use actual source directories and build commands, preserve handwritten instructions,
and assess only applicable host checks. Project-init keeps upstream-owned source
separate from its repository-owned Codex overlay; see
upstream synchronization.
Quality Gate
Content output follows the
content-review rubric
before publishing: score at least 85 on the standard scale, with the rubric's
Critical/Warning limits and format-specific scoring. Source validation and a build
are necessary evidence, not a substitute for the content review.
Local co-agent PR/push hooks and Kiro commit/push reviews are optional controls
with their own consent and failure policies. Their disabled or fail-open state
does not waive this repository's required GitHub checks. Read the data-sharing
disclosures before enabling
Kiro review/search delegation or
co-agent hooks.
Repository PRs require latest-HEAD AI Code Review and Codex package validation.
The configured review roster covers the full reviewed diff, and a chair verifies
findings. Active Critical/Major findings block merge. Missing, failed, or truncated
required coverage is an error, not an implicit pass. Review the HEAD, target branch,
and prerequisite PRs immediately before merge; retain the enforced diff caps.
Use current review policy,
CI review operations,
artifact review routing,
CI roster defaults,
AI review workflow, and
Codex validation workflow as the authority.
The privileged review executes trusted-base code and treats PR content as data;
the separate GitHub-hosted validation job checks the PR's generated packages without
provider secrets.
Project Structure
| Path | Maintained purpose |
|---|
plugins/ | Shared procedures, references, scripts, assets and source compatibility fixtures |
hosts/catalog.json | Shared plugin identity and version |
hosts/claude/, hosts/codex/ | Independent host execution settings |
packages/claude/, packages/codex/ | Generated self-contained installation packages |
.claude-plugin/marketplace.json | Claude marketplace |
.agents/plugins/marketplace.json | Codex marketplace |
scripts/sync-codex-plugins.py, scripts/codex/ | Host package assembler and Codex runtime templates |
scripts/pr-review/ | Repository CI review tooling and configuration |
doc-sites/ | Public documentation and frozen demo artifacts |
docs/ | Architecture, ADRs, runbooks, and internal references |
tests/ | Repository validation and regression checks |
tools/remarp-vscode/ | Remarp editor extension |
Internal documentation and this README remain English. Public guides use English
source and Korean locale content under ADR-023.
Requested deliverables can use another language; literal aliases, syntax tokens,
fixtures and frozen demo payloads remain data.
README.ko.md is a compatibility pointer to this README.
Release notes describe historical behavior, not current runtime guarantees.
Development
Edit maintained shared or host sources, then regenerate both host packages.
Do not repair generated copies alone. Run the repository checks from its root:
bash tests/run-all.sh
python3 scripts/test-plugins.py
python3 scripts/test-codex-plugins.py
python3 scripts/sync-codex-plugins.py --check
python3 scripts/eval-skills.py
Runtime/adapter changes also use the local probes described in
CLAUDE.md and the runtime verification guide. Those fixtures do not
certify external provider readiness. Public-site work uses the checks in
doc-sites/CLAUDE.md. See the
version tags for released artifacts.
