Skip to content
Musher Docs

Asset Types

Asset type vocabulary, naming conventions, and ecosystem mapping for bundle assets.

Overview

Asset types are free-form strings, not an enum. The asset_type field on BundleAsset accepts any string up to 50 characters. This design allows the type vocabulary to evolve as the agent ecosystem grows, without requiring schema migrations.

Recommended Types

Asset TypeDescriptionEcosystem Equivalent
promptSystem instructions, guidelines, or standalone prompt textCLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md
ruleModular rule files with optional path-scoping metadata.claude/rules/*.md, .cursor/rules/*.mdc, .roo/rules-*/*.md
skillComposable skill definitions (SKILL.md + supporting files).claude/skills/*/SKILL.md, Agent Skills standard
agent_specAgent or mode definitions with tool permissions and behavior.claude/agents/*.md, .agent.md, .roomodes
toolsetTool policies and MCP preset configurations.mcp.json, .cursor/mcp.json, .amazonq/mcp.json
templateParameterized templates for code generation or scaffoldingPrompt templates, .prompt.md files
configTool settings, environment config, or operational parameterssettings.json, config.toml, .aider.conf.yml
schemaJSON Schema, OpenAPI fragments, or data shape definitionsValidation schemas, API contracts

Naming Conventions

  • Use lowercase with hyphens for multi-word types (e.g. mcp-config, not mcpConfig or MCP_CONFIG).
  • Keep types short and descriptive — the 50-character limit is generous but brevity aids readability.
  • Prefer singular nouns (prompt, not prompts).

Ecosystem Context

The recommended types above are derived from a survey of ten major agent harnesses (Claude Code, Codex CLI, Cursor, Windsurf, Aider, GitHub Copilot, Cline, Roo Code, Amazon Q, Gemini CLI). Despite different naming conventions, the functional categories collapse to roughly the same set:

  • Instructions and prompts — every tool supports markdown-based instruction files.
  • Modular rules — 7 of 10 tools support multiple rule files loaded by directory scan.
  • Skills — the Agent Skills standard (SKILL.md) has reached 8 of 10 tools.
  • Agent definitions — role-based agent/mode configuration is supported by most tools.
  • MCP server configs — 9 of 10 tools support MCP, making it near-universal.

None of these tools provide built-in versioning for their configuration files. Bundles fill this gap by offering immutable, SHA-256-verified snapshots of asset collections.

Asset Support Matrix

How asset categories map across the ten surveyed agent harnesses. A tool name indicates documented support for that category.

CategoryBundle TypeSupported By
System instructionspromptAll 10 harnesses
Modular rulesruleClaude Code, Cursor, Windsurf, Roo Code, Cline, Amazon Q, GitHub Copilot
Skills / commandsskillClaude Code, Codex CLI, Cursor, GitHub Copilot, Cline, Roo Code, Windsurf, Gemini CLI
Agent / mode definitionsagent_specClaude Code, Codex CLI, Cursor, GitHub Copilot, Roo Code, Amazon Q, Gemini CLI
Tool policies & MCP presetstoolsetAll except Aider (9 of 10)
TemplatestemplateGitHub Copilot (.prompt.md), general-purpose
Settings / configconfigAll 10 harnesses (varying formats)
SchemasschemaGeneral-purpose (validation, API contracts)