Artifact Creator

Create the smallest complete artifact that satisfies the user's requested runtime and security boundary. Keep portable instructions separate from host packaging.

Overview

There is one portable Agent Skills contract, one open MCP protocol, and many different plugin, agent, hook, permission, and marketplace formats. This skill selects the correct authority before writing files and prevents host-specific details from leaking into the portable core.

When to use

Use for a new or substantially revised skill, subagent, plugin, MCP integration, hook, or marketplace entry. For a production modernization involving research, migration, security review, and release evidence, use production-upgrade instead. For review without changes, use artifact-validator.

Prerequisites

• The requested outcome and target repository or destination.
• Existing project instructions and the target host's current specification.
• Authorization for any local writes. Creation does not imply permission to install, publish, deploy, open a PR, or merge.

The pre-authorized Read, Write, and Edit capabilities apply only to local files inside the user-selected target. Every network, shell, installation, and external mutation capability remains behind host and project approval.

Instructions

1. Read project instructions and inspect the target tree before choosing a format. Preserve generated-file ownership and existing public identifiers.
2. Classify the artifact using references/artifact-contracts.md. If the request mixes a portable skill with host wiring, design the portable skill first and the host adapter second.
3. Establish the current authority:

• Agent Skill: the open Agent Skills specification plus any declared local overlay.
• MCP server: the Model Context Protocol specification and selected SDK.
• Plugin, subagent, hook, command, or catalog: the target host's current specification.
• This marketplace: repository instructions and its canonical validators.

1. Derive least privilege from the actual workflow. Do not copy a broad tool list from another artifact. Separate read-only analysis from mutations.
2. Create only useful files. A skill requires SKILL.md; add scripts when deterministic execution matters, references for conditional depth, assets for output material, and evals for behavior that needs proof.
3. For MCP work, define tools around bounded user outcomes. Validate every input, cap output and retries, redact secrets, separate reads from writes, and require explicit confirmation for high-impact mutations.
4. For host adapters, follow references/host-adapters.md. Do not claim another runtime is supported merely because it can read Markdown.
5. Validate with artifact-validator, run focused tests, and report unresolved assumptions and operations that still require approval.

Output

Return the created paths, the authority used for each artifact, validation commands and results, support limitations, and the next authorization boundary.

Error handling

• If the target host is unknown, create only the portable Agent Skill and leave host packaging unclaimed.
• If current official documentation conflicts with repository policy, stop and surface the conflict at the owning authority.
• If a secret, destructive default, unbounded retry, path escape, or unsupported runtime claim appears, fail closed and do not scaffold around it.
• If the destination already exists, preserve it and propose an update or migration; never overwrite without explicit authorization.

Examples

• "Create a skill that audits release readiness" produces one portable skill and only the resources required by its checks.
• "Build a Goose skill" produces a portable Agent Skill and verifies Goose's native path before offering installation instructions.
• "Create an MCP server and Claude plugin" designs the server against MCP and treats the Claude plugin configuration as a separate adapter.

Resources

• Artifact contracts
• Host adapter rules