Skip to content

SKILLS.md

Latest commit

 

History

History
73 lines (56 loc) · 2.53 KB

SKILLS.md

File metadata and controls

73 lines (56 loc) · 2.53 KB
Raw

Skills

A skill is a manifest that defines a specialized agent role: a preferred backend, a preferred model, and a system prompt. Wintermolt ships 79 built-in skills under skills/ and discovers user-installed ones at runtime from ~/.wintermolt/skills/.

Manifest format

Each skill is a directory containing one skill.json:

{
  "name": "code-reviewer",
  "description": "Reviews diffs for correctness, security, and clarity.",
  "backend": "claude",
  "model": "claude-sonnet-4-20250514",
  "role_prompt": "You are a senior reviewer. Flag bugs, security issues, and unclear naming. Cite line numbers. Suggest minimal patches.",
  "tools": ["file_read", "grep", "glob", "bash"]
}
Field Required Notes
name yes Unique identifier (kebab-case).
description yes One-line summary surfaced to the orchestrator agent.
backend yes One of ollama, claude, openai, deepseek, qwen, gemini, forai.
model yes Model tag for the chosen backend.
role_prompt yes System prompt injected when the skill is dispatched.
tools no Subset of built-in tools the skill may invoke. If omitted, all tools are available.

Built-in skills

Live in skills/<name>/skill.json. Examples:

skills/
├── code-reviewer/
├── 3d-animator/
├── audio-spatialize/
├── blender-rigger/
├── cli-anything-harness/
└── ... (79 total)

List from the REPL:

> /skills
> /skills code-reviewer

Or via the skills tool — the orchestrator agent can dispatch a skill when the user's request matches a skill's description.

Adding a custom skill

  1. Create ~/.wintermolt/skills/my-skill/skill.json.
  2. Restart Wintermolt (or invoke /skills reload).
  3. Reference it: > use my-skill to <task>.

Custom skills override built-ins of the same name.

Switching backends per skill

Every skill picks its own backend. The orchestrator may run on Ollama while a claude-grader skill runs on Claude — Wintermolt swaps backends transparently before dispatching, then restores the orchestrator's backend afterward.

If a skill names a backend you haven't configured (no API key, service down), Wintermolt falls back to the current backend and logs a warning to stderr.