Skill Creator V2

Skill Creator V2 is a production-grade meta-skill for creating, improving, evaluating, hardening, and packaging AI skills and skill groups.

It turns skill creation from an improvised writing task into a structured engineering workflow with explicit gates for scope, dependencies, failure modes, evals, evidence, packaging, and release readiness.

What It Does

• Creates single skills, skill groups, and orchestrator-worker systems.
• Separates visible skills from internal workers, reusable satellite skills, and shared modules.
• Improves existing skills while preserving behavior and regression evidence.
• Verifies tool, MCP, API, and package dependencies before relying on them.
• Blocks unsafe dependency assumptions such as unpinned latest installs.
• Runs lint, eval, benchmark, behavioral-eval, packaging, and final review helpers.
• Requires evidence before claiming that a skill is production-ready.

Who It Is For

• Codex and Claude Code users who maintain reusable local skills.
• AI workflow builders who need skill groups instead of one large prompt.
• Teams building agent systems with tools, MCPs, scripts, browser work, or external sources.
• Reviewers who need evidence-backed readiness reports instead of "looks good" claims.

Install

npm install -g skill-creator-v2

Install into one runtime:

skill-creator-v2 install --target codex
skill-creator-v2 install --target claude-code
skill-creator-v2 install --target gemini

Install into every supported local runtime root:

skill-creator-v2 install --all

Preview writes without changing files:

skill-creator-v2 install --all --dry-run

Supported targets:

• codex
• claude-code
• gemini
• antigravity
• vs-codium
• qwen
• zai-glm
• kimi
• github-copilot

Some runtimes do not have a confirmed native SKILL.md auto-loader. For QWEN, Kimi, and Z.AI/GLM, the installer creates a local skills/skill-creator-v2 folder plus an AGENTS.md projection that points the runtime/user to the skill.

Basic Usage

After installation, ask your agent:

Use skill-creator-v2.

Create a production-grade skill for <workflow>.
It should declare inputs/outputs, dependencies, failure modes, tests, evals, and a final readiness report.

For an existing skill:

Use skill-creator-v2.

Improve the existing skill at <path>. Preserve behavior, add production gates, run regression checks, and package the result separately.

For a skill group:

Use skill-creator-v2.

Create a skill group for UI intelligence: research references, extract patterns, synthesize design directions, and guard originality.

Hybrid Skill Groups

Skill Creator V2 does not treat every worker as a new installed skill. For groups, it now decides the runtime visibility of each role:

• merge_into_parent: a step that should stay inside the parent skill.
• internal_worker: a real worker with a WORKER.md contract, hidden inside one visible orchestrator.
• reusable_satellite_skill: a standalone skill that also participates in the group.
• shared_module: policy, schema, registry, template, or helper data that is not a skill.

This keeps runtime skills/ folders clean while preserving true reuse. A hybrid group can have one visible orchestrator, internal workers under group/workers/, reusable satellite skills, and shared modules under group/shared/.

What To Expect

The output should be more than a single Markdown prompt. A serious result usually includes:

• SKILL.md
• references/
• scripts/
• assets/
• agents/
• evals/
• deterministic validation commands
• a packaging artifact
• a readiness report with evidence and caveats

Why It Is Different

| Area | Skill Creator V2 | Typical hand-written skill | | --- | --- | --- | | Quality gate | Explicit production-ready gate and final review | Often subjective | | Dependencies | Declared, verified, allowlisted/blocklisted | Often implicit | | Failures | Closed taxonomy and retry policy | Often vague | | Evidence | Commands, artifacts, evals, reports | Often not captured | | Structure | Progressive disclosure | Often one flat prompt | | Skill groups | Orchestrator-worker contracts | Often unclear boundaries |

Case Study: UI Intelligence Group

The UI Intelligence group started as a set of design-research skills. Real runs showed that catalog pages and hero screenshots were not enough. The workflow was upgraded to require:

• live destination URLs from award directories
• 3-5 scroll-state screenshots where possible
• public DOM/CSS/font/script metadata
• effect cards that capture principles without copying code, layouts, images, text, SVGs, or unique animations
• originality review before any design handoff
• a feedback loop from failed runs back into the skills themselves

See case-studies/ui-intelligence.md.

Documentation

• English
• Russian
• French
• Spanish
• German
• Chinese
• Korean
• Japanese
• Runtime targets
• 10/10 rubric
• Eval runtime notes

Evidence Status

The package includes deterministic checks and eval fixtures. Current public evidence includes green GitHub Actions, unit tests, lint, taxonomy fixtures, meta-eval contract checks, 17 recorded independent behavioral executor outputs with 51/51 assertions passed, package inspection, temp-home installer smoke, npm-tarball installer smoke, and native CLI availability smoke for 7 locally available clients.

Important caveat: this proves the current public meta-eval corpus and package installer behavior. Future high-risk generated skills or skill groups still need their own behavioral eval evidence, and native skill auto-loading is adapter-dependent for some runtimes.