ワンクリックで
codex-second-opinion
Non-Codex agents only. Get an independent second opinion from Codex CLI on decisions with meaningful trade-offs.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Non-Codex agents only. Get an independent second opinion from Codex CLI on decisions with meaningful trade-offs.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Respond to unreplied review comments on the current GitHub Pull Request. For each comment, either commit a fix, push, and reply with the commit hash, or reply with concrete rationale for keeping the code as-is. Does not approve or request changes.
日本語の技術文書・書籍原稿の文章規範。整形(一文一行、引用ブロック、脚注、コラム記法)、段落と論証の構成(パラグラフライティング)、論証の厳密さ(ツッコミどころの除去)、読み手の負荷の管理、視点と語り、演出の抑制、LLM っぽい空句の禁止、冗長の排除を定める。日本語で技術書の章、草稿、記事、解説文を書くとき、または推敲・リライトするときに使用する。
Non-Antigravity agents only. Get an independent second opinion from Antigravity CLI (agy). agy cannot be driven headlessly from a sandboxed agent Bash, so this skill hands the user a ready-to-run command and folds in their pasted reply.
Methodology for iteratively improving agent-facing instructions (skills / slash commands / CLAUDE.md / code-gen prompts) via bias-free executor + two-sided evaluation (self-report + instruction-side metrics). Meta-skill, invoke ONLY when the user explicitly asks for an "empirical" eval of a prompt or skill, or for the Iter-0 description / body consistency check. Do NOT auto-invoke after every skill edit; this loop is operator-triggered by name.
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Draft a commit message from staged changes. Matches the style of recent commits in the repo, emphasizes why over what, and does not run `git commit`.
| name | codex-second-opinion |
| description | Non-Codex agents only. Get an independent second opinion from Codex CLI on decisions with meaningful trade-offs. |
If you are a Codex agent, STOP. Tell the user this skill is for non-Codex agents only.
$ARGUMENTS if provided; otherwise infer the question from the current task context. Keep it focused on decision-relevant information:You are providing a second opinion on a decision.
Context:
<context from the current task>
Question:
<the specific question or decision point>
Provide your independent analysis. If you disagree with the current approach, explain why and suggest alternatives.
sandbox_mode="danger-full-access" (nested sandbox-exec is prohibited on macOS):Pre-choose the output path yourself (used by both code branches below). Pick a unique literal filesystem path in advance — e.g., /tmp/codex-opinion-<unix-timestamp>-<random>.log, assembled in your working context (not via mktemp inside the snippet, because that binds the path to a shell variable you cannot retrieve later). Reuse the exact same literal path verbatim in the Bash invocation and in the follow-up Read. This avoids any dependency on a harness-specific output-capture tool (BashOutput, task output files, etc.).
If Bash supports run_in_background: true (typical case):
# Bash tool: run_in_background: true, timeout: 300000
# Substitute <TMPFILE> with the literal path you chose above.
codex -c 'sandbox_mode="danger-full-access"' exec - > <TMPFILE> 2>&1 <<'CODEX_PROMPT'
<constructed prompt>
CODEX_PROMPT
Wait for the background task to finish, then Read <TMPFILE>. Do NOT add sleep / kill -0 / pgrep poll loops — they burn tool-call budget and the harness's completion signal is authoritative. An interim Read "to peek at progress" is also unnecessary; the file is incomplete until Codex exits. If you are a nested subagent (running inside an Agent/Task tool), do NOT use this background path — a subagent that ends its turn while Codex is still running will lose the background task. Use the foreground fallback below instead. After reading, rm <TMPFILE>; if the harness denies rm, leave the file — /tmp is reclaimed by the OS. TaskOutput is for Agent/Task tool tasks — do not use it here.
Foreground fallback (use when run_in_background is unavailable, OR when you are a nested subagent):
# Bash tool: timeout: 300000
codex -c 'sandbox_mode="danger-full-access"' exec - > <TMPFILE> 2>&1 <<'CODEX_PROMPT'
<constructed prompt>
CODEX_PROMPT
Bash blocks until Codex finishes or the tool timeout fires. On return, Read <TMPFILE> and then rm <TMPFILE>. Do not use & + $! polling — it adds PID tracking for no benefit once you control the Bash tool timeout. Do NOT use trap 'rm -f $TMPFILE' EXIT in any snippet that backgrounds a process — the trap fires when the launching shell exits, deleting the file before the background process writes to it.
Parse output: ignore startup logs, MCP registration/errors, MCP tool-call traces (e.g. voicevox.text_to_speech, serena.find_symbol — any inline JSON or tool invocation lines from the caller's MCP context leaking into Codex's output, including cases where Codex itself invokes an MCP tool mid-reasoning), reasoning traces, and trailing metadata (tokens used, session id, token summary). The substantive answer is the main plain-text block; Codex sometimes prints it once inline and echoes it again after metadata. Tiebreaker when both copies are complete: use the post-metadata echo — it is Codex's canonical final form.
Present both perspectives (labeled by model), highlight agreements and disagreements. Evaluate Codex's suggestion critically — do not blindly adopt it.
On any failure — codex not found, auth error, network error, or timeout — report the reason to the user and proceed with your own analysis only. Discard partial output on timeout as unreliable.
Use a heredoc (<<'CODEX_PROMPT') to pass prompts via stdin. This avoids shell escaping and quoting issues. Codex exec - reads from stdin, so this is not subject to ARG_MAX.