Hm.
GitHub Get Hm
Docs / Agent

Instructions, prompts and skills

AGENTS.md, SYSTEM.md, prompt templates, skills, MCP servers and your own command tools.

Instruction files#

Hm reads the instruction files other agents use: AGENTS.md, or CLAUDE.md when there is none, from ~/.config/hm and from each folder between your home folder and the workspace, outer folders first.

A SYSTEM.md (.hm/SYSTEM.md, else ~/.config/hm/SYSTEM.md) replaces the base prompt, or adds to it when its first line is <!-- append -->.

The book icon in the session header and /context list the files a session uses. A session reads them once when it starts, so its prompt stays cached.

agent:
  context:
    enabled: true
    files: [AGENTS.md, CLAUDE.md]   # first one found per folder wins
    max_bytes: 65536                # larger files are skipped with a warning

Prompt templates#

Each .md file in ~/.config/hm/prompts, .hm/prompts, .claude/commands or .pi/prompts is a /name command in the message box, listed in its completions with its description. Claude Code commands work as they are.

---
description: Review the staged change
argument-hint: "[focus]"
mode: rubber-duck
---
Review the staged diff. Focus on $ARGUMENTS.

@CONTRIBUTING.md
  • $ARGUMENTS is what you type after the command; $1, $2 … its words (quotes group them).
  • A line @path inlines a workspace file (guard.yaml rules apply).
  • Front matter can set description, argument-hint, mode and model.
  • Built-in commands win over a template of the same name.

Skills#

Skills are read from ~/.config/hm/skills, .hm/skills, .cursor/skills and .claude/skills.

MCP servers#

tools:
  mcp:
    github:
      transport: stdio            # stdio | http
      command: github-mcp-server
      args: [stdio]
      env: { GITHUB_TOKEN: "${GITHUB_TOKEN}" }
      approval: ask               # auto | ask | deny, for all its tools
    docs:
      transport: http
      url: https://mcp.example.com/mcp
      headers: { Authorization: "Bearer ${DOCS_TOKEN}" }

Command tools#

Turn any command-line program into a tool. Parameters are passed as separate arguments, never through a shell, so a value can't inject another command.

tools:
  commands:
    run_migrations:
      description: Apply database migrations up to a version
      command: ./scripts/migrate up {version}
      timeout: 2m
      approval: ask
      params:
        version: { type: string, description: Target version, required: true }
tools:
  web:
    search: { provider: duckduckgo }   # duckduckgo | searxng | brave | tavily | exa
Edit this page on GitHub
esc