| name | skill-building |
| description | Design and create agent skill packages — instruction-based and MCP-based skills, manifest format |
Skill Building
This skill teaches you how to create Markus skill packages — directory-based artifacts that teach agents new capabilities. A skill can work in two ways (or both):
- Instruction-based: A
SKILL.md file with instructions injected into the agent's context, guiding it to use existing tools in new ways.
- MCP-based: A bundled MCP server (script) that provides entirely new tools to the agent. The skill directory can contain any scripts, config files, or resources the MCP server needs.
Most skills are instruction-based. Use MCP-based skills when the capability requires new tools that don't exist yet (e.g., connecting to an external API, browser automation, hardware control).
Artifact Directory
CRITICAL: Skill artifacts MUST be saved under this exact path — the Builder page, install system, and deliverable detection all depend on it:
~/.markus/builder-artifacts/skills/{skill-name}/
├── skill.json # Manifest (auto-created from your JSON output)
├── SKILL.md # Instruction document (you write via file_write)
├── README.md # Human-readable documentation (optional)
└── ... # Any other files: scripts, MCP servers, configs, templates, etc.
Do NOT write artifacts to ~/.markus/shared/, your working directory, or any other location. Only ~/.markus/builder-artifacts/skills/ is recognized by the system.
A skill directory can contain any files needed for the skill to work — not just SKILL.md and README.md. For example:
- MCP server scripts (e.g.,
server.mjs) that provide new tools to the agent
- Configuration templates or data files
- Helper scripts used by the instructions
When the user installs the artifact, the entire directory (all files) is deployed to ~/.markus/skills/{skill-name}/. SKILL.md is loaded and injected into the agent's context when the skill is activated. If the manifest declares mcpServers, those servers are started and their tools registered to the agent. skill.json contains metadata used by the skill registry. README.md provides documentation for humans browsing or sharing the skill.
Two-Step Workflow
Output the skill in two steps — manifest first, then content files. Never put file content inline in the JSON.
Chat Mode vs Task Mode
- Chat mode (user conversation): Output the manifest JSON in a ```json code block → system auto-saves and creates the directory → then use
file_write for each content file.
- Task mode (assigned task): Use
file_write to write the manifest JSON file directly (e.g., file_write("~/.markus/builder-artifacts/skills/{name}/skill.json", ...)) → then use file_write for each content file. When submitting deliverables, set the reference to the artifact directory path.
- A2A mode (agent-to-agent): Same as task mode — write all files via
file_write.
Step 1: Output Manifest JSON
In chat mode: Output the skill configuration as a JSON code block. The system auto-saves it.
In task/A2A mode: Write the manifest JSON file directly via file_write.
This JSON contains ONLY metadata — no file content.
Instruction-based skill (most common):
{
"type": "skill",
"name": "skill-name-kebab-case",
"displayName": "Skill Name",
"version": "1.0.0",
"description": "When and why an agent should use this skill",
"author": "Your Name",
"category": "custom",
"tags": ["tag1", "tag2"],
"skill": {
"skillFile": "SKILL.md"
}
}
MCP-based skill (provides new tools via a bundled server script):
{
"type": "skill",
"name": "my-api-connector",
"displayName": "My API Connector",
"version": "1.0.0",
"description": "Connect to My API for data retrieval and actions",
"author": "Your Name",
"category": "custom",
"tags": ["api", "connector"],
"skill": {
"skillFile": "SKILL.md",
"requiredPermissions": ["network"],
"mcpServers": {
"my-api": {
"command": "node",
"args": ["${SKILL_DIR}/server.mjs"]
}
}
}
}
Notes on MCP servers:
${SKILL_DIR} is resolved at load time to the skill's actual directory path — use it to reference bundled scripts.
- The
command can be any executable (node, python3, npx, etc.).
- The MCP server communicates via JSON-RPC 2.0 over stdio (stdin/stdout).
- Tool names exposed by MCP servers are automatically prefixed with the server name (e.g.,
my-api__tool_name). Mention these prefixed names in SKILL.md.
- You can also use externally published MCP servers:
"command": "npx", "args": ["-y", "some-mcp-server@latest"].
The system automatically saves this JSON and creates the directory. After that, you proceed to write files.
Step 2: Write Files with file_write
After the JSON is saved, write each file individually using file_write. The base path is ~/.markus/builder-artifacts/skills/{skill-name}/ (use the name from your JSON).
Write files in this order:
-
SKILL.md (REQUIRED) — The instruction document with YAML frontmatter and comprehensive Markdown body:
- YAML frontmatter with
name and description (must match manifest)
- Overview of what the skill does
- Step-by-step instructions referencing actual tools (or MCP tool names if MCP-based)
- Error handling guidance
- Examples of typical input/output
-
MCP server script (if MCP-based) — e.g., server.mjs implementing the MCP protocol over stdio. Must handle initialize, tools/list, and tools/call JSON-RPC methods.
-
README.md (optional) — Human-readable documentation for browsing or sharing.
-
Any other files — Helper scripts, templates, config files, data files, etc.
Example file_write calls:
file_write("~/.markus/builder-artifacts/skills/git-changelog/SKILL.md", "---\nname: git-changelog\ndescription: Generate changelogs from git history\n---\n\n# Git Changelog\n\n## Overview\n...\n\n## Instructions\n...\n\n## Examples\n...")
file_write("~/.markus/builder-artifacts/skills/git-changelog/README.md", "# Git Changelog\n\nA skill that helps agents generate changelogs from git history...\n")
Field Reference
Top-level fields
type: Always "skill"
name: MUST be English kebab-case (e.g., git-changelog, web-scraper). Must match SKILL.md frontmatter name. This is the directory name and identifier.
displayName: Human-readable skill name, can be in any language
version: Semver (default "1.0.0")
description: When and why an agent should use this skill (can be in any language)
category: Typically "custom" for user-created skills
tags: Array of descriptive tags
skill section (REQUIRED)
skillFile: Always "SKILL.md" — the entry point instruction document
requiredPermissions: (optional) Array of permissions: "shell", "file", "network", "browser"
mcpServers: (optional) Map of MCP server name → config. Each config has command, args?, env?. Use ${SKILL_DIR} in args/env to reference the skill directory.
alwaysOn: (optional, boolean) If true, the skill's instructions are automatically injected into every agent's context at startup. Only use this for foundational skills that all agents must always follow (e.g., self-evolution). Default is false — non-alwaysOn built-in skills are listed as "available" in the agent's identity section and can be activated on-demand via discover_tools.
After Creation
CRITICAL: Creating an artifact is NOT the same as installing/deploying it. Creating writes files to builder-artifacts/; installing makes the skill available to agents in the live org. NEVER auto-install. Only install when the user explicitly asks. This applies to ALL modes (chat, task, A2A).
Once all files are written, tell the user:
- The skill has been created and saved — summarize what was created (name, purpose, what agents can do with it).
- Ready to install — the user can install from the Builder page, or ask you to install it (you would use
package_install). Do NOT install unless asked.
- To modify or improve this skill (e.g., add more instructions, update examples, fix edge cases), just continue the conversation here — describe what you want to change and I'll update the files directly.
Guidelines
- Instructions in SKILL.md should reference actual tools:
shell_execute, file_read, file_write, file_edit, grep_search, glob_find, list_directory, web_fetch, web_search, gui — or MCP tool names if the skill provides its own tools
- For MCP-based skills, document every tool with its prefixed name (e.g.,
my-api__search) in SKILL.md so the agent knows how to use them
- Be specific — include actual CLI commands, file paths, and URL patterns
- Include error handling: what to do when commands fail, pages don't load, etc.
- Provide examples of typical input/output for each workflow step
- Skills should be self-contained: an agent reading the instructions should know exactly what to do
- Consider composability: skills that work well alongside other skills
- After outputting the JSON, immediately proceed to write files via
file_write — announce what you're writing
- When creating MCP server scripts, use only Node.js built-in modules (no npm dependencies) for maximum portability, or use
npx to reference published packages
Rules
- DO NOT use names that conflict with built-in skills. Check the dynamic context for existing skill names.
- DO NOT put file content in the JSON. Always use
file_write for files.
- DO NOT write artifacts to
~/.markus/shared/ or your working directory. Always use ~/.markus/builder-artifacts/skills/{name}/.
- The
name field MUST be English kebab-case (e.g., git-changelog, not 网页抓取器). This is the directory name and package identifier.
- The
name field and SKILL.md frontmatter name must match exactly.
- All top-level fields must be the correct type:
author must be a plain string (your name, e.g. "John") — NOT an object. tags must be an array of strings. version must be semver string. description must be a string. The system validates the manifest on write and will reject malformed files.