skill-author

skill-author — draft a new chatbot skill

Generate a fresh SKILL.md under skills-dev/<name>/ so the user can start iterating immediately. The output must satisfy GA's SKILL.md schema so the chatbot's FileBasedSkillsProvider picks it up via live-reload, and Claude Code can also load it as a plugin skill.

Inputs the user provides

• A short name (kebab-case, like progression-mood or chord-info).
• A one-paragraph description of what the skill should do.
• Optional: example user prompts that should trigger it.
• Optional: classification — catalog (deterministic lookup, no LLM per call), computation (calls an MCP tool deterministically), or hybrid (LLM synthesizes over a deterministic substrate).

Pre-flight checks (refuse if these fail)

1. No name collision: glob skills/<name>/SKILL.md and skills-dev/<name>/SKILL.md. If either exists, refuse and tell the user to either pick a different name OR explicitly invoke the graduator/redrafter pattern (move existing to drafts first).
2. Name is kebab-case: ^[a-z][a-z0-9-]*$. Reject snake_case, PascalCase, spaces.
3. Description is non-empty and contains at least one usage hint like "use when" or "when the user asks". The description is what the orchestrator's intent router (and Claude Code's slash-command matcher) score against — vague descriptions misroute.

Output template

---
name: <name>
description: |
  <one paragraph: what the skill returns, when to use it, what
   evidence kind it produces. Include 2-3 user-prompt fragments
   the orchestrator should match against.>
triggers:
  - "<trigger phrase 1>"
  - "<trigger phrase 2>"
  - "<trigger phrase 3>"
license: internal
compatibility:
  agent-framework: ">=1.0.0-preview"
metadata:
  authoring-style: <deterministic-catalog | mcp-tool-driven | hybrid-llm>
  origin: "drafted by skill-author agent on <ISO date>"
  evidence-kinds:
    - <catalog_lookup | mcp_tool_call | llm_synthesis>
---

# <Skill title>

## When the chatbot dispatches to this skill

<Concrete user-prompt examples — at least 3. The orchestrator's intent
router compares user messages against THIS body as well as the
description, so concrete examples here improve routing precision.>

## What the skill returns

<Output structure: sections, fields, formats. Include a short example
output if the format is non-obvious.>

## How the skill computes

<Reference the implementation: an `IOrchestratorSkill` class for
catalog skills, an MCP tool name for tool-driven skills, or "LLM
synthesis grounded on <X>" for hybrid skills. State whether the
behaviour is fully deterministic or whether the LLM is in the path.>

## Refuse to invent

<List the things the skill MUST NOT hallucinate — chord voicings
outside the catalog, scale memberships, set-class numbers, etc. The
chatbot's groundedness validator gates on these.>