Claude-Codex Skills Marketplace
A compact marketplace of standalone engineering skills for Claude Code and Codex covering review, audit, optimization, testing, product discovery, architecture, and safe publishing.
Evidence is limited to README and two CI workflows; the actual SKILL.md contents were not provided. The maintainer suite (62/63/64) commits, pushes, publishes releases and Discussions, but no file-level evidence of confirmation gates or approval gates is shown — deducted. Least-privilege CI permissions (contents: read) are a positive signal but concern the repo's own CI, not installed-skill permissions. The validator is pinned (@2.1.207); otherwise dependency security is unevidenced. No rollback documentation. Author named in LICENSE, but publisher identity is unverified.
README statements align with the repo layout and the CI validates both the Claude marketplace and repository contract scripts, giving some self-consistency (2); actual SKILL.md files are absent from evidence, so failure messaging can only be credited as described (1). Dual-host packaging manifests are declared for availability (2).
Explicit dual-host support (Claude Code and Codex) with separate install instructions and a clear audience (2); per-skill trigger conditions live in files not provided, so trigger precision rests on the README's 'explicit loading trigger' claim (1); no statement on other environments.
Excellent information architecture: indexed skill numbering, directory tree, canonical layout and shared SKILL tree explained (3); complete, executable install notes for both hosts (3); naming rules with canonical allocation documented (3); full MIT license (3). Deductions: no FAQ or concrete output examples (1), no versioning/changelog (1), known limitations limited to one static-validation caveat (1), maintenance responsibility rests on a single unverified individual without governance (2).
Output contracts (fixed report order, 200-line ceiling, evidence accounting) are clearly described (2); structured decision gates and evidence checks add real value over unguided prompting (2); opt-in suite installation, no MCP/runtime dependencies, on-demand reference loading keep cost light (2). None of this is execution-verified, so full marks are not available.
Claims are checkable via scripts/validate-repository.ps1 and CI, and the README's tree and paths are traceable (2); compliance claims against external specs cannot be cross-corroborated from the provided evidence (1); the README explicitly separates static validation from behavioral validation — good fact/inference separation (2).
- This is a static review; no skill was executed, and the actual SKILL.md contents, triggers, and safety gates remain unverified.
- The maintainer suite (62–64) can commit, push, and publish; review its confirmation and rollback logic before installing, and avoid installing it unless needed.
- Publisher identity is unverified; marketplace installation implies trust in future updates — consider pinning or reviewing changes.
- No version numbers or changelog exist, making behavioral drift across skills hard to track.
- CI dependencies are version-pinned, but runtime dependency availability of downstream skills is unverified.
What does this agent do, and when should you use it?
This repository is a compact plugin marketplace providing 24 standalone engineering skills for the Claude Code and Codex hosts, organized into seven suites: a Review Suite (plan and delivery review), a Codebase Audit Suite (documentation, code health, test suite, architecture, and persistence auditors), an Optimization Suite (performance, dependency upgrades, modernization, benchmark comparison, surgical changes), a Testing Suite, a Product Discovery Suite, a Maintainer Suite, and an Architecture Suite. Each skill is a plugin directory with a SKILL.md containing domain checklists, evidence requirements, and a final self-check; skills have a 200-line ceiling and load references only for the relevant workflow branch. The repo ships a portable Agent Plugins v1 manifest, host-specific catalogs for Claude Code (.claude-plugin/marketplace.) and Codex (.agents/plugins/marketplace.), and a static catalog site. It deliberately contains no MCP servers, orchestration hierarchy, or evaluation harness, and users install only the suites they need.
Skills run as instructions inside the host (Claude Code or Codex): Plan Reviewer validates an implementation plan against repository evidence before execution; Delivery Reviewer reviews delivery acceptance, regressions, and release risks; the five auditors assess documentation trust, cross-cutting code health, test value and isolation, system boundaries and dependency topology, and query/transaction consistency; Performance Optimizer profiles and experiments to keep only measured improvements, while Benchmark Comparator compares alternatives under a frozen, reproducible experiment contract; the Architecture Suite (ln-71 through ln-76) produces baselines, current-state documents, target design proposals, decision records, diagrams, and migration plans from repository evidence; and the Maintainer Suite validates, commits, pushes, and remotely verifies repository changes, then publishes GitHub releases and Discussions announcements. Skills are invoked by namespaced names, e.g. /review-suite:ln-12-delivery-reviewer.
- A team lead validates an implementation plan against repository evidence with ln-11-plan-reviewer before letting an agent execute it.
- An engineer inheriting a legacy codebase runs the audit suite (ln-21 to ln-25) to assess documentation, code health, tests, architecture, and persistence.
- A performance-sensitive project uses ln-31-performance-optimizer and ln-34-benchmark-comparator for reproducible experiments, keeping only measured wins.
- A team upgrading dependencies without a big-bang rewrite uses ln-32-dependency-upgrader for reversible, verified batches.
- An architect uses the architecture suite to generate current-state documents, decision records, and migration plans from existing repo evidence.
- A skill-repo maintainer uses the maintainer suite (ln-61 to ln-64) to review skill content and publish GitHub releases and Discussions announcements.
What are this agent's strengths and limitations?
- Works with both Claude Code and Codex hosts and follows the portable Agent Plugins v1 standard, with one canonical copy of each skill (skills/<name>/SKILL.md).
- Every skill carries explicit evidence checks, applicability reasons, and a final self-check; reports follow a uniform structure (result, scope, evidence, verification, completion).
- Seven independently installable suites let you adopt only what you need; the repo is intentionally minimal with no MCP or orchestration dependencies.
- Includes a static validation script, scripts/validate-repository.ps1, covering structure, skill contracts, catalogs, and documentation coverage.
- Depends on host CLIs (Claude Code or Codex); it cannot run outside those environments and offers no MCP or standalone runtime path.
- The README states static validation establishes format and consistency only and does not certify task performance, so real skill quality must be verified by the user.
- No independent evaluation harness exists to quantify skill outcomes, and additional repository guidance (AGENTS.md) was not available for deeper review.
- Contributors must learn the numbering scheme (1x–7x), SKILL_TEMPLATE.md authoring format, and AGENTS.md validation rules before adding or editing skills.
How do you install or deploy this agent?
In Claude Code, add the marketplace and install only the suites you need:
/plugin marketplace add levnikolaevich/claude-code-skills
/plugin install review-suite@levnikolaevich-skills-marketplace
/plugin install codebase-audit-suite@levnikolaevich-skills-marketplace
/plugin install optimization-suite@levnikolaevich-skills-marketplace
/plugin install testing-suite@levnikolaevich-skills-marketplace
/plugin install product-discovery-suite@levnikolaevich-skills-marketplace
/plugin install maintainer-suite@levnikolaevich-skills-marketplace
/plugin install architecture-suite@levnikolaevich-skills-marketplace
/reload-pluginsIn Codex:
codex plugin marketplace add levnikolaevich/claude-code-skills
codex plugin add review-suite@levnikolaevich-skills-marketplace(repeat with the other suite names.)
For local development, load one plugin directly: claude --plugin-dir ./plugins/review-suiteHow do you use this agent?
After installation, invoke a skill by its namespaced name, e.g. /review-suite:ln-12-delivery-reviewer for delivery review, /codebase-audit-suite:ln-22-codebase-auditor for a codebase health audit, or /architecture-suite:ln-76-architecture-migration-planner for a migration plan. Each skill loads its skill-local references only for the relevant workflow branch and produces reports in a consistent order: result, scope, evidence, verification, completion. No MCP server or shared runtime is required.
How does this agent compare with similar options?
The README positions the repo as a compatible implementation of Claude Code's native plugin mechanism and the OpenAI plugin guide, aligned with the Agent Plugins v1, Agent Skills, and Claude Code skill/plugin specifications; no third-party competitors are named in the source.