Skill

Tools for exploring, creating, updating, and reviewing agent skills.

The Skill toolkit gives you a structured workflow for building your own Claude Code / Agent Skills compatible skills: research what already exists before building, scaffold a new skill from a template, iterate on it, and run quality reviews before shipping.

Installation

Claude Code:

npx skills add cloudvoyant/codevoyant

OpenCode / VS Code Copilot: See the installation guide.

Typical Workflow

skill:explore → skill:new → skill:update → skill:review
(research)       (create)    (refine)       (quality gate)

Build a New Skill from Scratch

/skill explore "linear integration"   # Research existing skills first
/skill new linear-push                # Scaffold from template
/skill update linear-push             # Refine any dimensions
/skill review linear-push             # Audit spec compliance + conventions

Improve an Existing Skill

/skill update dev-commit              # Choose dimensions to improve
/skill review dev-commit --effectiveness  # Full effectiveness audit

Skills

Explore Existing Skills

Research available skills before building something new:

/skill explore                            # Interactive topic prompt
/skill explore "linear integration"       # Explore a specific topic
/skill explore "ci monitoring" --thorough # Deeper scan with more sources

Runs parallel research agents against npx skills find, agentskill.sh, and published skill repos. For each result, shows:

• Source URL and install command
• SKILL.md excerpt (frontmatter + first 50 lines)
• Absorb assessment — whether you can extend rather than build from scratch

Output lands in .codevoyant/explore/{slug}/ and can be passed directly to /skill new via --research.

Flags:

• --thorough — deeper scan, more sources, higher haiku agent count

Create a Skill

Scaffold a new skill from the canonical template:

/skill new                                 # Interactive naming
/skill new linear-push                     # Named skill
/skill new linear-push --research .codevoyant/explore/linear/explore.md

Guides you through:

1. Loading research context (from --research or .codevoyant/explore/)
2. Confirming the high-level skill overview
3. Designing the step-by-step workflow
4. Selecting destination directory (skills/ for public, .claude/skills/ for private)
5. Writing SKILL.md, LICENSE.md, and any references/ files
6. Running mise run skills:validate to confirm it passes

Output is committed to skills/{skill-name}/ (public) or .claude/skills/{skill-name}/ (private).

Flags:

• --research <path> — seed the session with an existing explore artifact

Update a Skill

Improve an in-progress skill plan or an existing committed skill:

/skill update                         # Auto-detect active skill plan
/skill update dev-commit              # Target specific skill or plan
/skill update dev-commit > add --dry-run flag   # Minor annotation
/skill update dev-commit >> restructure agent steps  # Major annotation

Two modes:

• In-progress plan — apply > (minor) or >> (major) inline annotations to the active plan.md
• Existing skill — choose which dimensions to improve: prompt quality, performance, agent additions, or skill dependencies

Changes are always planned and confirmed before being written. Runs mise run skills:validate after applying.

Annotation syntax:

• > note — minor clarification on a step; applies inline without structural change
• >> note — significant change (scope, step removal/addition); requires confirmation

Review a Skill

Audit a skill for spec compliance, claudevoyant conventions, and effectiveness:

/skill review                             # Review most recently updated skill
/skill review dev-commit                  # Review specific skill by name
/skill review path/to/SKILL.md            # Review by path
/skill review skills/                     # Review all skills in a directory
/skill review dev-commit --effectiveness  # Include effectiveness dimensions

Checks:

Spec compliance (Step 2):

• Frontmatter completeness: name, description, license, compatibility
• Portability tier accuracy (Tier 1 / 2 / 2b / 3 based on features used)
• Claudevoyant conventions: pushy description, argument-hint, disable-model-invocation, Step 0 argument parsing, no markdown tables, references/ for large content, > **Compatibility**: note

Effectiveness review (Step 3, optional — pass --effectiveness to run automatically):

Rates five dimensions as Strong / Adequate / Weak:

• Objective Clarity — purpose is unambiguous from the description alone
• Instruction Quality — steps are a concrete recipe with verifiable completion conditions
• Template / Artifact Quality — output templates follow no-tables rule and produce useful artifacts
• Trigger Accuracy — description trigger keywords match intended invocation patterns
• Actionability — a first-time user can run the skill without reading external docs

Produces a review report at .codevoyant/review/{skill-name}.md. Auto-fixes mechanical issues (missing test commands, blank sections) and flags judgment calls one at a time.

Flags:

• --effectiveness — include the effectiveness review pass automatically

List All Commands

/skill help                 # Not yet implemented — see this page