ycc Bundle Author

This skill scaffolds new source-of-truth surfaces under ycc/ — skills and optional agents. Skills themselves provide slash invocation. It is a contributor workflow scoped strictly to this repo's ycc/ tree. It writes nothing under .cursor-plugin/ or .codex-plugin/ (those are regenerated), and it emits the exact follow-up commands needed to sync the derived bundles after scaffolding completes.

Arguments

Parse $ARGUMENTS:

• skill-name (required, first positional) — kebab-case identifier, no leading or trailing hyphens.
• --skill-only — stamp user-invocable: false into skill frontmatter for passive guidance that should not appear in the slash menu. Model invocation remains available. Otherwise the skill is user-invocable as a native slash command.
• --with-agent — also scaffold ycc/agents/<skill-name>.md.
• --dry-run — preview only; writes nothing.

The older --with-command flag is accepted as a legacy no-op. It never creates a wrapper or overrides --skill-only.

Questions: class clarify (Phase 0 anti-pattern match "proceed?"): headless (no question tool or chat turn), print the matched anti-pattern and stop; scaffold nothing. See ${CLAUDE_PLUGIN_ROOT}/skills/_shared/references/question-policy.md.

Phase 0: Guards

Before scaffolding, verify all of the following. If any check fails, abort immediately with a clear, specific message.

1. skill-name matches the regex ^[a-z][a-z0-9]*(-[a-z0-9]+)*$. Reject names with uppercase letters, leading/trailing hyphens, or consecutive hyphens.
2. None of the target paths already exist:
   • ycc/skills/<skill-name>/
   • If --with-agent: ycc/agents/<skill-name>.md Use test -e for each path and abort if any returns true.
3. Name does not collide with an existing skill or agent. Use ls on ycc/skills/ and ycc/agents/ to confirm.
4. Re-read ${CLAUDE_PLUGIN_ROOT}/skills/bundle-author/references/when-not-to-scaffold.md and check whether any anti-pattern applies to the user's request. If a match is detected, surface it explicitly and ask whether to proceed before continuing.

Phase 1: Preview

Read ${CLAUDE_PLUGIN_ROOT}/skills/bundle-author/references/surface-map.md. Present the user with the full list of files that will be created:

• ycc/skills/<skill-name>/SKILL.md (with user-invocable: false stamped in the frontmatter when --skill-only is set)
• ycc/skills/<skill-name>/references/ (directory, no files created by default)
• If --with-agent: ycc/agents/<skill-name>.md

Note that scripts/ subdirectories are not created by default. If the skill will need bash helpers, the contributor adds ycc/skills/<skill-name>/scripts/ manually after reviewing the scaffolded output.

Phase 2: Dry-run check

If --dry-run was passed: print the file list from Phase 1, then print the exact post-scaffold commands from Phase 4 (substituting the real skill name), and STOP. Do not invoke the helper script.

Phase 3: Scaffold

Invoke the helper:

${CLAUDE_PLUGIN_ROOT}/skills/bundle-author/scripts/scaffold-skill.sh <skill-name> [--skill-only] [--with-agent]

The helper copies templates from ${CLAUDE_PLUGIN_ROOT}/skills/bundle-author/references/templates/ and substitutes {{NAME}} and {{DESCRIPTION}} placeholders (references/templates/skill-template.md, run when scaffolding any skill in Step 1). --with-agent also renders references/templates/agent-template.md; read it only to check the rendered agent file. If the helper exits with a non-zero status, STOP and surface its stderr verbatim — do not attempt to repair or work around the failure.

Phase 4: Post-scaffold steps

Always emit the following block to the user after a successful scaffold, substituting <skill-name> with the actual name. Omit line 2 unless --with-agent was passed.

Next steps:
1. Edit ycc/skills/<skill-name>/SKILL.md to replace TODOs.
2. If --with-agent: edit ycc/agents/<skill-name>.md.
3. If you added scripts under ycc/skills/<skill-name>/scripts/: chmod +x them.
4. Regenerate derived bundles: ./scripts/sync.sh
5. Validate: ./scripts/validate.sh

Evals

A new skill ships with 3 eval cases under ycc/evals/: two prompts that must load the skill and one near-miss that must load a sibling or nothing (under trigger/), each graded with tool_used: Skill. Manual-only skills (disable-model-invocation: true) replace the two positive cases with no-autoinvoke cases under ycc/evals/policy/ (min: 0, max: 0). Skills with a hard gate (confirmation, dry-run, no-write default) also get a case under ycc/evals/behavior/ with deterministic graders. Copy an existing case, run it with ./scripts/eval.sh --case <name>, and keep descriptions fixed until it passes. See ycc/evals/README.md.

Important Notes

• Never writes under .cursor-plugin/ or .codex-plugin/ — those are regenerated by ./scripts/sync.sh from sources in ycc/.
• Never edits ycc/.claude-plugin/plugin.json or .claude-plugin/marketplace.json — those are maintained by the bundle-release workflow.
• Refuses generic scaffolding for external repos — see ${CLAUDE_PLUGIN_ROOT}/skills/bundle-author/references/when-not-to-scaffold.md.
• Prefer extending an existing skill over adding a near-duplicate. Run ls ycc/skills/ to confirm no existing skill covers the same domain before proceeding.