菜单
Use when the user asks to plan a task before coding (e.g. "plan-loop", "/plan-loop", "plan this out first"). Draft a detailed plan, cross-review it with Claude, triage skeptically, iterate, then gate for sign-off.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use when the user asks to take a task end-to-end autonomously (e.g. "autodev", "/autodev", "build this on autopilot"). Runs the full chain — plan → implement → review → PR → converge — stopping only at real decisions.
Use when the user asks for a cross-review by Claude (e.g. types "claude-review", "/claude-review", or "have Claude review this"). Invokes Claude Code headlessly to review the current branch/changes, prints Claude's findings verbatim, then triages them skeptically instead of auto-agreeing.
Use when the user wants the current code proactively reviewed and fixed (e.g. "review-loop", "/review-loop", "review what I have and fix it"). Self + cross-model review → triage → fix → update PR, loop until clean.
Use when the user wants a product/design/UX critique (e.g. "product-review", "/product-review", "is this the right thing to build / the right UX?"). Questions what to build and how users will perceive it, not just code correctness.
| name | plan-loop |
| description | Use when the user asks to plan a task before coding (e.g. "plan-loop", "/plan-loop", "plan this out first"). Draft a detailed plan, cross-review it with Claude, triage skeptically, iterate, then gate for sign-off. |
| metadata | {"short-description":"Plan → cross-review with Claude → triage → gate"} |
Settle a plan before any code, Codex-side. Same workflow as Claude's
/plan-loop.
Canonical procedure: read ~/.codex/skills/skyhook-skills-commands/plan-loop.md and follow it
(detailed plan with file:line evidence + Certain/Confident/Needs-input tiers,
cross-review, triage, iterate, gate). Codex translations:
claude -p "Critique this implementation plan as a skeptical senior engineer. Find flawed assumptions, missing steps, hidden complexity, simpler alternatives, and risks. Challenge it; don't rewrite it. PLAN:\n\n<the full plan>" --model opus --permission-mode default --allowedTools Read Grep Glob Bash
claude -p needs Keychain access — run it
escalated/unsandboxed (see the claude-review skill's auth note), or it
falsely reports "Not logged in" from inside Codex's sandbox.--auto resolves
Needs-input with best judgment, logs each call to NOTES.md, and proceeds —
except a ceiling decision (prod/deploy/data-migrations · external-publish/money
· security/auth/secrets · breaking-API/destructive-file-ops) stops regardless.Narrate phases with the glyph-tagged banners and close with the short summary ledger defined in the canonical file. When invoked by the autodev skill, the approved/assumed plan flows straight into implementation.