prompt-pilot

name	prompt-pilot
description	Optimize a rough prompt with an independent subagent, present several CONCISE ready-to-run variants for the user to pick or edit, then EXECUTE the chosen prompt directly in the current session. Unlike prompt-optimizer (advisory only — it just hands back text), prompt-pilot closes the loop: optimize → choose → run. TRIGGER when the user wants both optimization AND execution: "优化并执行", "优化提示词后直接做", "优化指令然后执行", "帮我把这个需求优化好再做", "optimize and run my prompt", "improve this prompt then do it", "refine my prompt and execute it", or when the user explicitly invokes /prompt-pilot. DO NOT TRIGGER when the user only wants advice/text back with no execution ("优化prompt", "改进prompt", "rewrite this prompt", "how to write a prompt for") — that is prompt-optimizer. DO NOT TRIGGER for "优化代码" / "优化性能" / "optimize this code" / "optimize performance" — those are refactoring or performance tasks. DO NOT TRIGGER when the user says "直接做" / "just do it" with no request to optimize — execute the task normally without this skill.
origin	fork
metadata	{"author":"oh-my-claudecode","version":"1.3.0","changelog":"1.3.0 — UI/interaction/motion prompts now use the project-local routing\n contract: broad Web polish starts with impeccable, motion work is\n delegated to the curated Emil specialists, and final appearance is\n checked with visual-verdict. Expo and SwiftUI use stack-specific\n substitutions instead of receiving the Web chain blindly.\n1.2.0 — Swept Medium/Low review items: terminal behavior for a repeated\n clarifying-questions loop (no infinite re-dispatch); tightened the\n over-broad \"When to Use\" line so a bare vague task can't trigger it;\n read-only subagent is now tool-enforced (prefer analyst/Explore, not\n general-purpose) and glm-5.2 for cost; variants must match the\n raw prompt's language; \"Other\" disambiguates custom-prompt vs\n reject/re-optimize; added explicit non-trigger examples.\n1.1.0 — Hardened the optimize→execute loop after an adversarial review:\n (1) forward pasted conversation context/attachments into the\n optimizer subagent (it cannot see the thread); (2) defensive JSON\n parsing — strip fences, retry once, fall back to raw text / original\n prompt instead of fabricating variants; (3) grounding discipline so\n the subagent never bakes unverified paths into an auto-executed\n variant; (4) explicit, mandatory execution gate with a\n destructive-operation hard stop.\n1.0.0 — Initial release.\n"}
argument-hint	<your rough prompt / task idea>
user-invocable	true
disable-model-invocation	false
model	glm-5.2

Prompt Pilot

Take a user's rough prompt, hand the optimization work to an independent subagent, get back 3 concise, ready-to-run prompt variants, let the user pick / edit / supplement one, then run the chosen prompt in this session.

This is the execution sibling of prompt-optimizer. Where prompt-optimizer stops at "here is a better prompt", prompt-pilot continues all the way to "...and here is the result of running it".

When to Use

• User wants the prompt improved and then executed: "优化并执行：…", "帮我把这个需求优化好再做", "optimize and run my prompt"
• User pastes a vague task and explicitly asks to sharpen it before doing it — there must be an optimize-and-run signal, not merely a vague request. A bare vague task with no such signal is a normal task, not this skill.
• User explicitly invokes /prompt-pilot

Do Not Use When

• User only wants the optimized text handed back → use prompt-optimizer
• User says "直接做" / "just do it" with no optimize request → execute normally
• "优化代码" / "优化性能" / "optimize this code/performance" → refactor/perf task

Workflow

Run these four steps in order. Do NOT do the optimization yourself in the main context — that is the subagent's job (keeps the main context clean and gives an independent pass).

Step 1 — Capture the raw prompt

Take everything after the trigger as the raw prompt. If the user only typed the trigger with no task, ask once: "What do you want to get done?" Then proceed.

Step 2 — Dispatch ONE independent optimizer subagent

Dispatch a read-only subagent with the instruction block below — prefer one with no edit access (analyst, or Explore/explore) so the read-only rule is tool-enforced, not just prompt-level. Avoid general-purpose here: it can mutate files. GLM-5.2 is plenty for this optimization pass — and for the actual execution in Step 4. The subagent runs a compressed version of the prompt-optimizer methodology (intent + gap analysis, light project probe, ECC component match) but its deliverable is 3 concise variants, NOT a giant essay.

Send the subagent exactly this (fill in <RAW_PROMPT>, <CWD>, and <CONTEXT>). Critical — the subagent cannot see this conversation; it only gets what you paste. Before dispatching, gather anything in the thread that bears on the task — pasted error logs, stack traces, file snippets, the described content of screenshots, and any clarifications the user already gave — and put it in . If there is genuinely none, set to . Skipping this is the skill's biggest quality leak: bug reports usually carry their diagnosis in the pasted payload.