Guide for writing or refining prompts for Claude or GPT-5.6 Sol, distilled from each vendor's official best practices. Use for new prompts, debugging an existing prompt, or "prompt engineering", "system prompt", "Claude/GPT/OpenAI prompt". Accepts arg: "claude" or "gpt".
The ONLY way to call GPT (a.k.a. Codex), Grok, GLM, Kimi, DeepSeek, or MiMo. Use whenever the user wants to ask, delegate to, or get a second opinion from GPT, Grok, GLM, Kimi, DeepSeek, or MiMo. Do NOT run the codex, grok, glm, kimi (km), deepseek (ds), or mimo (mm) CLI directly — from the main agent or a subagent; always use this skill's relay call command. Triggers on "ask/have/send to/get/delegate to gpt/codex" or the same with "grok"/"glm"/"kimi"/"deepseek"/"mimo", "second opinion", "relay".
Dispatch multiple independent agents to answer the SAME complete question from different analytical lenses, then synthesize. Reach for this by default, without waiting to be asked, when redundant cross-model judgment could change the decision: ambiguous architecture/design tradeoffs, high-stakes or hard-to-reverse changes, competing root-cause hypotheses, or failure-mode-sensitive reviews — and whenever you would otherwise spawn 2+ independent reviewers for one question. Skip for trivial lookups, deterministic transforms, routine small edits, mechanical syncs, and single-correct-answer tasks. With no leading number, autonomously decide N and gpt-pro from the question every time (N=1 is the bottom-rung anchor, never a lazy default); with an explicit number (e.g. `prism 2 1`), honor it verbatim and skip auto-sizing. There is no reasoning-effort knob. Scale above the anchor only when decision risk justifies the 8N-agent cost.
Drive a goal through review→fix iteration by composing goal-elicit, goal-drive, and multi-model review (prism, Claude-only). A thin, modeless, stepped loop — one phase per invocation (elicit → spec-review → drive → review → you pick which fixes apply → fix → re-review) — with state on disk so it resumes after interruption or compaction. It auto-handles the ~83% of out-of-scope/unmapped findings (you never see them) and surfaces only the small actionable batch to confirm — never blind-applies (unsafe). Use for "loop this to done with review", "close the review-fix loop", "goal-loop". Review needs Claude (prism), degrades off-Claude. Interactive by default; **`--auto`** runs it unattended/headless under Claude Code's native `/goal` — both gates become fail-closed policies (oracle-gated safe-subset auto-fix + a deferred morning queue), never blind-apply. Does NOT interview (that's goal-elicit). Skip for one-off edits and lone reviews.
Send a prompt to ChatGPT Pro Extended via gpt-pro-relay on macmini — over SSH from any other machine, or directly when invoked on macmini itself. Use for "ask gpt-pro", "send to gpt-pro", "use gpt-pro", "Pro Extended take", "ask the deep model", or "second opinion from chatgpt pro". The wrapper (not the caller) polls through flaky networks; the agent fires one backgrounded call and waits for the notification.
Audit Codex/OpenClaw skills: loaded roots, duplicate skills, unused skills, prompt-budget costs, compact descriptions.
Read arxiv papers via their TeX source for full-fidelity math, tables, and figures. Use whenever the user shares an arxiv URL or paper ID and wants to discuss or understand the paper — even without "read". Triggers on arxiv URLs (abs, pdf, html), bare IDs like "2401.12345", or "explain this paper". Do NOT use for note generation (use note-gen).
Autonomous experiment loop faithful to Karpathy's autoresearch. Set up a branch, establish a baseline, then loop forever: edit, commit, run, measure, keep or revert. Use when asked to "run autoresearch", "optimize X in a loop", or "start experiments".