Scaffold AgentKit skills, commands, agents, and rules

VerifiedSafe

Scaffolds a new AgentKit asset (skill, command, agent, rule, or guideline) with correct frontmatter and location.

Sby Skills Guide Bot
DevelopmentBeginner
1010/18/2026
Claude CodeCursor
#scaffolding#agentkit#claude-code#templates#frontmatter

Recommended for

Our review

Scaffolds a new AgentKit asset (skill, command, agent, Cursor rule, or guideline) at the correct location with proper frontmatter, based on the provided templates.

Strengths

  • Automatically detects context (AgentKit checkout vs. consuming project with .claude/) and picks the right destination path
  • Prevents overwrites by checking for collisions before writing
  • Validates that the description is trigger-shaped rather than a summary
  • Reminds about post-creation best practices: no secrets, body length limit, and referencing shared/guidelines

Limitations

  • Assumes the templates (SKILL.template.md, etc.) are reachable in templates/ or the AgentKit install
  • For Cursor rules, requires the user to supply the NN prefix and globs array themselves
  • The guideline kind only works inside the AgentKit repo, not a consuming project
When to use it

Use when you want to quickly create a new skill, command, agent, or rule with a compliant structure and frontmatter.

When not to use it

Avoid for editing or refactoring existing assets, or when AgentKit templates aren't available in the environment.

Security analysis

Safe
Quality score86/100

The skill scaffolds new AgentKit assets by creating directories and files, reading templates, and checking for collisions. It uses Write/Edit/Read/Glob/Bash only for legitimate file creation and path inspection, with no destructive commands, network exfiltration, secret handling, or obfuscation.

No concerns found

Examples

Nouveau skill
/scaffold skill pdf-extractor Use when the user asks to extract text or tables from a PDF file.
Nouvel agent
/scaffold agent backend-engineer Reviews Go and Python backend services for API design and error handling.
Nouvelle règle Cursor
/scaffold rule react-components Use when editing React component files to enforce hooks and prop-typing conventions.

description: Scaffold a new skill, command, agent, or Cursor rule with correct frontmatter and location. argument-hint: <kind> <name> [description] allowed-tools: Bash, Read, Write, Edit, Glob

Scaffold a new AgentKit asset. Parse $ARGUMENTS as: <kind> <name> [description].

  • <kind> = skill | command | agent | rule | guideline
  • <name> = kebab-case (my-skill, backend-engineer, go-service)
  • [description] = optional one-liner; ask for it if missing

Where to create it

First determine context: are we inside an AgentKit checkout (has skills/, commands/, agents/ at the root), or inside a consuming project with .claude/ installed?

| Context | Path | | --- | --- | | Consuming project (.claude/ present) | .claude/skills/<name>/SKILL.md etc. | | AgentKit library root | skills/<name>/SKILL.md etc. |

Full dispatch table (consuming project):

| kind | Creates | | --- | --- | | skill | .claude/skills/<name>/SKILL.md | | command | .claude/commands/<name>.md | | agent | .claude/agents/<name>.md | | rule | .cursor/rules/<NN>-<name>.mdc (ask for NN prefix) | | guideline | shared/guidelines/<name>.md (AgentKit only) |

Before writing

  1. Check for collisions. If the target path exists, stop — don't overwrite.
  2. Validate the description. Must be trigger-shaped: "Use when the user asks to…". Reject summary-shaped descriptions like "Handles X-related tasks."
  3. For rules, ask for the globs array — don't guess the file pattern.

Use the template

Find the right template from templates/ (or from where AgentKit is installed):

  • Skill → SKILL.template.md
  • Command → command.template.md
  • Agent → agent.template.md
  • Rule → rule.template.mdc

Fill every <placeholder>. Mark genuinely optional sections with <!-- TODO -->.

After creating

  1. Print what was created and what still needs filling in (the <...> fields).
  2. Remind the user:
    • Description must be trigger-shaped or discovery won't work.
    • No project-specific names, credentials, or internal URLs in assets meant to be reused.
    • Reference shared/guidelines/ for security, multi-cloud, or document rules rather than duplicating them inline.
    • Keep the body under ~120 lines — it's loaded into context on every use.
Related skills