create-project-skill

name	create-project-skill
description	Scaffold a project-local skill under a consuming repo's `.agents/skills` tree for Codex and Claude by default, with Codex-only and bridge-only modes.

Create Project Skill

Contract

Prereqs:

• Run inside the target consuming project's git work tree.
• The request is for a project-local skill, not a repo-owned runtime-kit managed skill. Use create-skill for agent-runtime-kit managed skills.
• The caller has supplied or accepted a lowercase hyphenated skill name that starts with project- (project-local skills) or private- (skills under the private overlay home, $AGENT_PRIVATE_SKILLS_HOME).
• Existing files are never overwritten without explicit user approval.
• Claude-only creation is unsupported. Claude exposure is a bridge to the canonical .agents/skills source tree, not a second source location.

Inputs:

• Skill name, description, intended entrypoint, and whether the skill needs skill-owned scripts/, tests/, references/, fixtures/, or bin/.
• Creation mode:
  • default / --target both: create the canonical skill source and ensure .claude/skills -> ../.agents/skills.
  • --codex-only / --target codex: create only the canonical skill source and do not mutate .claude/.
  • --bridge-only: create or verify only the Claude bridge for an existing .agents/skills tree.
• Optional .agents/scripts/<command>.sh thin wrapper for project slash-command style dispatch.
• Optional project validation command to run after scaffolding.

Outputs:

• .agents/skills/<skill>/SKILL.md.
• Optional .agents/skills/<skill>/scripts/<skill>.sh executable stub.
• Optional .agents/skills/<skill>/{tests,references,fixtures,bin}/ support directories.
• Default .claude/skills symlink and .gitignore entry for .claude/ unless --codex-only is used.
• Optional .agents/scripts/<command>.sh wrapper that delegates to the skill-owned script or another canonical project command.
• A validation summary with created paths and any skipped optional surfaces.

Failure modes:

• The current directory is not inside a git work tree.
• The skill name is malformed (does not match ^(project|private)-[a-z0-9-]+), ambiguous, or collides with an existing .agents/skills/<skill> path.
• A requested wrapper path already exists and replacement was not approved.
• --target claude, --claude-only, or removed --link-only is requested.
• --bridge-only is requested without an existing .agents/skills tree.
• The workflow would mutate runtime-kit manifests, product render output, golden snapshots, or global runtime homes.
• Validation fails after scaffolding.

Entrypoint

Use the bundled helper for deterministic file creation:

$HOME/.hermes/plugins/meta/skills/create-project-skill/scripts/create-project-skill.sh <skill-name> --description "One-line description."

Default creation exposes the skill to both Codex-style .agents/skills consumers and Claude through the project-local bridge. Use Codex-only only when the current repo should not receive a .claude/ local directory:

$HOME/.hermes/plugins/meta/skills/create-project-skill/scripts/create-project-skill.sh <skill-name> --codex-only

Use bridge-only when .agents/skills already exists and only Claude discovery needs to be wired:

$HOME/.hermes/plugins/meta/skills/create-project-skill/scripts/create-project-skill.sh --bridge-only

Optional support surfaces are explicit:

$HOME/.hermes/plugins/meta/skills/create-project-skill/scripts/create-project-skill.sh <skill-name> --with-script --with-tests --with-wrapper <command>

Workflow

1. Resolve the target project root with git rev-parse --show-toplevel.
2. Confirm this is a project-local skill request. If the user wants a runtime-kit managed skill, switch to create-skill.
3. Normalize and validate the skill name as lowercase hyphenated text that starts with project- (project-local) or private- (private overlay); reject any other prefix.
4. Inspect .agents/skills/, .agents/scripts/, .claude/, project AGENTS.md / CLAUDE.md, and nearby local skill conventions before editing.
5. Refuse to overwrite existing skill directories, scripts, wrappers, symlinks, or project policy files unless the user explicitly approves replacement.
6. Create SKILL.md with front matter, H1, Contract, Entrypoint or Scripts, Workflow, and Boundary sections.
7. Add only the support folders explicitly requested. Keep stubs small and mark TODO behavior clearly when implementation remains.
8. In the default mode, create or verify .claude/skills -> ../.agents/skills and ensure .claude/ is ignored. Skip all .claude/ mutation only when --codex-only is requested.
9. If a slash-command style wrapper is requested, create a thin .agents/scripts/<command>.sh wrapper that delegates to the skill-owned script or another canonical project command.
10. Create .agents/scripts/pre-pr.sh only when --with-pre-pr-stub is requested.
11. Refuse removed Claude-only flags: --target claude, --claude-only, and --link-only.
12. Run focused validation: test -f, shell syntax for generated scripts, project-owned checks when available, and the user's requested validation command.
13. Report created paths, skipped optional paths, validation status, and next implementation work.

Boundary

This skill owns project-local skill scaffolding only. It does not mutate manifests/skills.yaml, manifests/plugins.yaml, targets/, build/, tests/golden/, sandbox expected skill lists, or live global runtime homes. If project-local creation needs a reusable dry-run/apply planner, reference graph, or machine-readable mutation contract, extract that deterministic behavior to released nils-cli first and then call it from this workflow.