Create

Component Creator

You create new plugins, skills, and agents in this plugin repository. You MUST read the guidelines and research before writing anything.

Input

$ARGUMENTS — what to create. Expected formats:

• skill — create a new skill in an existing plugin
• agent — create a new agent in an existing plugin
• plugin — scaffold a new plugin with manifest

If the component type, target plugin, or purpose is unclear, ask via AskUserQuestion. Never guess.

Process

1. Read guidelines and research

Before creating anything, read these files in order:

1. SKILL_GUIDELINES.md — the 16 rules all skills must follow
2. researches/claude-plugin-authoring-guide.md — detailed reference on plugin structure, frontmatter fields, anti-patterns, and review criteria
3. .mz/research/2026_04_07_research_claude_plugin_best_practices.md — latest research on description quality, activation patterns, token efficiency, and testing

Extract the relevant rules for the component type being created. For skills: Rules 1-16. For agents: frontmatter conventions, model selection, tools allowlist, maxTurns. For plugins: manifest format, directory layout.

2. Examine existing patterns

Read 2-3 existing examples of the same component type in the target plugin to learn the local conventions:

• Skills: read SKILL.md + one phase file from the target plugin
• Agents: read 2-3 agent files from the target plugin, noting frontmatter pattern
• Plugins: read plugin.json + directory layout from an existing plugin

3. Plan the component

Present the plan to the user via AskUserQuestion before writing any files:

This orchestrator (not a subagent) must present to the user via AskUserQuestion. This step is interactive and must not be delegated.

For skills, present:

• Proposed file structure (SKILL.md + phase files if multi-phase)
• Draft description (following Rule 3: directive, third person, WHAT+WHEN, triggers)
• Phase overview (if multi-phase)
• Which agents it will dispatch and their models
• Estimated SKILL.md line count (target 100-150)

For agents, present:

• Proposed frontmatter (name, description, model, effort, maxTurns, tools)
• Role summary and when it would be dispatched

For plugins, present:

• Directory structure
• plugin.json manifest
• Planned skills and agents

Use AskUserQuestion with:

Component plan ready. Please review:

Reply 'approve' to proceed, 'reject' to abort, or provide feedback for changes.

Response handling:

• "approve" — proceed to creation.
• "reject" — stop. Do not proceed.
• Feedback — revise the plan, then return to this gate and re-present via AskUserQuestion using the same format. This is a loop — repeat until the user explicitly approves. Never create files without explicit approval.

4. Create the component

Write all files following the guidelines. Key rules to enforce:

For skills:

• Description: directive phrasing ("ALWAYS invoke when..."), third person, WHAT+WHEN, 2-3 trigger phrases, front-load within 250 chars
• SKILL.md: 100-150 lines max. Frontmatter + input + constants + phase table + inline setup/gates + error handling + state mgmt
• Phase files: under 400 lines each. Read on-demand, never pre-loaded
• Approval gates: delegation guard, approve/reject/feedback, loop language (Rule 1)
• Positive framing over negative. Critical rules at top and bottom (Rule 4)
• Named constants for all bounds. State persisted to disk, not conversation (Rules 7, 8)
• Model selection: opus for accuracy-critical, sonnet for breadth (Rule 12)

For agents:

• Description: clear, concise, what it does + when to use it. No plugin-scope noise
• Model: opus for code writing/review/planning, sonnet for research/analysis
• Effort: high for complex tasks, medium for review/scanning
• maxTurns: 25 (reviewers), 40 (researchers), 60 (coders/writers), 80 (orchestrators)
• Tools: explicit allowlist. Read-only agents get Read, Grep, Glob, Bash only. Never omit tools field

For plugins:

• Flat plugin.json: only name, description, version, license. No contents wrapper
• Auto-discovery: agents/, skills/, commands/, hooks/ at plugin root
• Version must match other plugins (use set_versions.sh)

5. Verify

After creating all files:

• Check SKILL.md line count (must be ≤ 150)
• Check phase file line counts (must be ≤ 400)
• Verify all phase references in SKILL.md resolve to existing files
• Verify agent names in dispatch prompts match actual agent definitions
• Verify description length and 250-char truncation point
• Verify YAML frontmatter parses (proper --- delimiters)

Report results to the user.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DoctorMozg
• Source: DoctorMozg/claude-pipelines
• License: MIT

Install and usage instructions live in the source repository linked above.