ll-meta-skill-creator

name

ll-meta-skill-creator

description

Create or update LEE Lite workflow skills that keep a standard agent-skill shell while adding LEE governance files for contracts, structural and semantic validation, execution and supervision evidence, revision limits, and freeze gates. Use when an agent needs to scaffold or revise a governed workflow skill, especially for requests mentioning LEE workflow skills, ll.contract.yaml, ll.lifecycle.yaml, contract/evidence/supervisor/freeze, or workflows such as src-to-epic and epic-to-feat.

LL Meta Skill Creator

Overview

Create governed workflow skills for LEE Lite. Each generated skill must remain a standard skill folder that can run in compatible agent-skill shells, but it must add a LEE governance pack for contracts, validation, evidence, supervisor review, and freeze control.

Workflow

1. Capture the workflow boundary before writing files.
   • Record the workflow key, input artifact type, output artifact type, authoritative upstream ref, runtime mode, direct entrypoint, and freeze expectations.
   • Classify the request against E:\ai\LEE-Lite-skill-first\ssot\adr\ADR-038-运行时核心抽象边界与对象分层基线.MD first: primary abstraction, secondary abstraction, authority, carrier, and not-equal boundary.
   • Scaffold a governed workflow skill only when the primary abstraction is really Skill or Workflow; if the request is mainly Command, Tool, Task, Gate, or Artifact work, redirect it to the proper carrier or SSOT object instead of generating a new skill bundle.
   • Stop if the request does not define the workflow boundary well enough to write input and output contracts.
2. Scaffold the governed skill.
   • Run python scripts/init_lee_workflow_skill.py <skill-name> --path <dir> --input-artifact <type> --output-artifact <type>.
   • For LEE Lite repositories, keep the default --runtime-mode lite_native. Use --runtime-mode legacy_lee only when the user explicitly wants interop with an existing lee runtime.
   • Prefer the generated layout over ad hoc file creation so every workflow skill starts from the same LL baseline.
3. Fill the governance pack, not just SKILL.md.
   • Keep the generated SKILL.md as the standard-compatible shell.
   • Put workflow rules, contracts, lifecycle, evidence schemas, and gates into ll.contract.yaml, ll.lifecycle.yaml, input/, output/, evidence/, agents/, and workflow scripts.
4. Separate execution from supervision.
   • Executor content belongs in agents/executor.md, execution evidence, draft generation, and structural validation.
   • Supervisor content belongs in agents/supervisor.md, semantic review, revision decisions, and freeze approval or rejection.
5. Validate before handing the skill over.
   • Run python scripts/validate_lee_workflow_skill.py <path/to/generated-skill>.
   • If the standard skill validator is available in the environment, run it too.
6. Install when needed.
   • Use ll-skill-install to install a canonical skill into Codex or Claude Code as a workspace-bound adapter.
   • scripts/install_profile.py is a thin compatibility wrapper that delegates to ll-skill-install; do not invoke it directly unless maintaining legacy code.

Non-Negotiable Rules

• Keep the skill shell compatible with standard skills. Do not move LL governance fields into YAML frontmatter.
• Do not use a workflow skill scaffold to smuggle in a Command, Tool, Task, Gate, or session residue requirement.
• Require both input and output contracts. A workflow skill is incomplete if it only templates the output.
• Split structural validation from semantic validation. Structural checks should be scriptable; semantic checks should be reviewable and evidenced.
• In LEE Lite, do not point a newly generated workflow back to lee run unless the user explicitly asks for a legacy bridge.
• Do not let the executor issue the final semantic pass on its own output.
• Require execution evidence and supervision evidence before any freeze gate can pass.
• Keep lifecycle semantics lightweight. Use states to align review and evidence, not to rebuild a heavy workflow engine.

Frontmatter Guidance

• Default to standard name and description only.
• Add shell-specific extensions such as allowed-tools, context, or agent only when the target runtime and validator explicitly support them.
• Put LL-specific fields in ll.contract.yaml, never in frontmatter.

Scripts

• scripts/init_lee_workflow_skill.py
  • Creates a governed workflow skill scaffold with contracts, evidence schemas, agents, resources, and either a lite-native runtime stub or a legacy lee placeholder profile.
• scripts/validate_lee_workflow_skill.py
  • Validates that a governed workflow skill includes the required LL files and core contract fields.
• scripts/install_profile.py
  • Compatibility wrapper. --profile codex delegates to ll-skill-install; --profile standard copies the canonical skill without adapter rewriting.

References

• Read E:\ai\LEE-Lite-skill-first\ssot\adr\ADR-038-运行时核心抽象边界与对象分层基线.MD first. It is the project-wide runtime abstraction baseline for classifying Command, Skill, Tool, Agent, Task, Workflow, Session, Artifact, Evidence, and Gate.
• Read references/ll-governance-pack.md for the required file tree and file responsibilities.
• Read references/authoring-patterns.md when customizing contracts, semantic checklists, or supervisor rules for a concrete workflow such as src-to-epic.
• Read references/lite-native-guardrails.md when the target repository should run the skill directly in Codex or Claude Code instead of delegating to lee run.