Skill Generator

Entry Boundary

This is a direct entry skill, alongside execute and execute-one. It is not a Reactant tag skill or document router. Activate it only from an explicit user request to create a Reactant project-local skill, such as:

/skill-generator a translate skill that preserves citations and LaTeX commands

There is no inline form for this entry skill. execute and execute-one must never select or dispatch an apparent generator annotation in an authored document. Do not read, consume, or edit a document merely to obtain the generator request.

The generated directory is a standard Agent Skill. The router agent discovers the new worker by reading project skill roots on its next invocation; there is no registry and no programmatic tag scanner.

Input

Accept the user's direct natural-language description of the new skill and the current project root. If the intended behavior, name, or file effects are materially ambiguous, clarify them before creating files.

Workflow

1. Read the description and references/skill-contract.md in this skill directory. Infer the requested behavior without reducing the description to a machine tag grammar.
2. Choose a name made of lowercase letters or numbers separated by single hyphens. It must not be a reserved name (execute, execute-one, skill-generator, chat-user, resolve, chat-model, comment) or collide with another skill unless the user explicitly requests an allowed override.
3. Inspect the project-local skill root and create .reactant/skills/ if it does not exist.
4. Write .reactant/skills/<name>/SKILL.md according to the contract: YAML frontmatter with a matching name, a description explicitly limiting activation to router-dispatched tags, an Activation Boundary section, tag examples, input behavior, workflow, scope-modifier and context-reference handling, consumption behavior, and clean failure behavior. Add supporting scripts/, references/, or assets only when the skill needs them; keep the skill self-contained.
5. Validate the directory by reading it against the structural contract: frontmatter, naming, required behavior, self-containment, and reserved-name rules. If an optional structural Agent Skills or Reactant validator is available, it may be run as an additional check. Such a validator must not scan authored documents or classify annotations.
6. Fix any structural errors and validate again. If the skill cannot be made valid, remove only the incomplete directory created by this invocation and report the diagnostics.
7. Report the created path and the generated worker's annotation form. Do not modify an authored document as part of generation.

Rules

• Generated skills are tag-only workers; this generator does not create new entry skills.
• Validation is structural only. Reactant does not semantically verify what a generated skill will produce at runtime.
• Phase 1 skill generation and routing require no Python runtime or vendored reactant_core.py.
• On failure, do not leave an incomplete generated directory. Preserve all pre-existing project files and report the problem.