Claude Skills by kimmingul

Headless CLI orchestration for multi-LLM advisors (xllm). Use when you want authentic responses from external CLIs: claude, codex, gemini, antigravity (agy), grok, cursor, or local ollama/lmstudio/lemonade. Routes a task through the chosen provider's headless mode and saves a structured artifact. Primary building block for /xllm.

Setup and doctor for xllm. Creates artifact dirs, remembers advisor script path for this project, checks cloud/local providers, and prints recommended commands.

Multi-LLM review and synthesis for xllm (headless). Decomposes the task, runs 2+ external CLIs (codex, antigravity, grok, claude, ollama, lmstudio, …) via the advisor script, captures artifacts, then synthesizes agreements and disagreements. Primary multi-advisor skill (replaces the former OMC-style /ccg name).

Get one authentic external-model opinion from another vendor's CLI (codex, claude, gemini, grok, cursor, antigravity) or a local model (ollama, lmstudio, lemonade) without leaving your session. Use when the user wants a second opinion, cross-vendor review, or a local-LLM answer — e.g. "ask codex", "what does gemini think", "get a second opinion on this design". Advisors run read-only by default.

Delegate a bounded implementation task to another vendor's CLI executing in an ISOLATED clone — it edits, builds, and tests there, then hands back a verified git ref plus evidence; the user's checkout is never touched. Use when the user wants a different vendor (codex) to implement, fix, or refactor something, or wants work done off the main session. Escalation ladder: ask (opinion) → propose (static diff) → exec (verified branch).

Cross-vendor deliberation over a question or a code diff, four modes in one noun: roles (parallel advisors + host synthesis), blind (measured independent panel), debate (adversarial refutation), council (blind→debate pipeline). Use for "get multiple opinions", cross-vendor design or security review, "review this diff/branch with other vendors", or when a decision is consequential and claims need stress-testing. Advisors run read-only by default; nothing is auto-applied.

Route mechanical git prose — commit messages, PR titles/bodies, release notes, changelog entries — to the cheapest healthy model (free local ollama first) instead of burning main-session tokens. Use whenever the user asks to commit, open a PR, cut a release, or write a changelog and the prose can be drafted from the diff/log. The advisor writes text only (read-only); YOU run the git/gh commands.

Diagnose and configure xllm cross-vendor advisors for THIS project: scan the machine for installed advisor CLIs and local models, then run a short Q&A to pin which provider+model+effort each role should use here, with recommended defaults. Use for "set up xllm", "xllm doctor", "configure advisors for this project", or when the ask or review skills cannot find providers.