dzhng/skills

Selection

Select when

• You want a reusable Claude Code skill library rather than a new coding-agent host.
• A large feature should be decomposed into independently buildable and verifiable slices.
• The implementation plan must evolve when new facts emerge during coding.
• You want agent-made architectural and product choices recorded for later review.
• You need a structured cleanup, code-review, and documentation closeout pass.
• You want screenshot comparison and independent visual critique integrated into delivery.
• You work on WebGPU, three.js, TSL, or WGSL rendering and need detailed validation rules.
• You want to author or blind-evaluate additional Agent Skills.
• You explicitly want Claude Code or Codex used as a separately verified second agent.
• You prefer selective installation of small, editable Markdown skills.

Boundaries

Avoid when

• You need a fully autonomous result without human acceptance criteria or final review.
• You require every listed skill to be bundled by the supplied Claude plugin manifest.
• You cannot permit skills or agents to inspect and potentially modify a working tree.
• You lack the external Claude Code or Codex CLI required by their respective delegation skills.
• You require platform-independent GUI image review rather than the macOS Preview workflow.
• You need a proven security audit, compatibility matrix, or independently benchmarked reliability.
• You need a standalone application, MCP server, or hosted orchestration service.
• You cannot isolate broad-permission agent runs in dedicated worktrees.
• You need guaranteed multi-agent compatibility beyond repository claims.
• You require stable performance evidence without project-specific hardware and browser verification.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only selected first-party files were supplied, so many of the 22 listed skills, any helper scripts, assets, package configuration, tests, CI workflows, and marketplace metadata were not inspected.
• Uncertainty: repository creation and activity dates are in 2026 relative to the analysis context, so recency, adoption, and maturity signals cannot be independently reconciled.
• Uncertainty: the README claims compatibility with more than 70 harnesses, but no host matrix or cross-host test evidence was supplied.
• The Claude plugin manifest names only 12 skills, while the README enumerates 22, so plugin installation may not expose the full advertised collection.
• The full loop invokes a /goal capability that is not present in the supplied skill inventory or plugin manifest, suggesting an external harness dependency that is not fully documented here.
• The cited 1-day-16-hour unattended Codex run is repository-reported proof from one run and does not establish repeatable correctness or general reliability.
• No minimum Claude Code version, plugin update behavior, automated compatibility results, security policy, threat model, telemetry disclosure, or independent audit was supplied.
• Several workflows rely on external CLIs, authenticated network services, browsers, GPU hardware, project tests, or other environment-specific tools.
• The Codex and Claude delegation guidance includes broad-permission modes for some tasks; even with documented isolation controls, these modes materially increase local execution risk.
• Visual verification depends on saved artifacts and human judgment, while preview-shots is specifically limited to macOS Preview.
• Multi-hour or multi-day autonomous runs can consume substantial model budget and still require several hours of human work at the beginning and end according to the README.
• The renderer skill contains detailed, potentially version-sensitive claims about three.js, TSL, Playwright, Chromium, and GPU behavior that were not independently tested.

The skills operate inside development workspaces and can direct shell commands, local authenticated agent CLIs, browsers, and file edits. The Codex and Claude files describe broad-permission execution only for explicit delegation and pair it with dedicated-worktree isolation and mandatory diff review, but these are procedural controls rather than a sandbox guarantee. No security audit or threat model was supplied.