菜单
Start a new session. Trigger with /start at the beginning of a conversation. Reads the session brief, checks git status, summarizes where things stand, and increments the session number.
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。
基于 SOC 职业分类
/brief - Write session brief, commit, and push. Run at end of session to checkpoint all work.
Get an outside-Anthropic second opinion on any piece of work by sending it to Google Gemini with an adversarial critique prompt. Use when the user wants to pressure-test a decision, plan, assumption, blog draft, landing page, positioning statement, campaign, or chunk of code with a non-Claude model. Trigger when the user says 'devil's advocate,' 'play devil's advocate,' 'poke holes,' 'stress test this,' 'second opinion,' 'outside perspective,' 'non-Anthropic read,' 'sanity check from another model,' 'challenge this,' 'critique this from outside,' or '/devils-advocate.' This is explicitly a cross-model sanity check — do NOT substitute your own critique; the whole point is to surface blind spots Claude might share.
Analyze the current conversation for corrections, approvals, and patterns, then propose updates to the skill files that were used. Trigger with /reflect or automatically at end of /brief. Use when the user says 'reflect,' 'learn from this,' 'update the skill,' or 'remember this for next time.'
Audit all available Claude skills across user and project scopes — inventory them, map routing and overlap, score each on efficiency, reliability, clarity, maintainability, learning, currency, and safety, then produce a prioritized report and draft patches. Use when the user asks to audit, review, improve, optimize, modernize, clean up, consolidate, benchmark, or evaluate skills; to find outdated, duplicate, or inefficient skills; to identify missing skills; or to check whether skills follow current best practices. Diagnosis only — never edits a skill without explicit approval.
Sync Claude Code skills to Codex. Handles project-level (.claude/skills → .agents/skills) and user-level (~/.claude/skills → ~/.codex/skills). Trigger with /sync-skills or when the user mentions 'sync skills,' 'codex skills,' or 'sync to codex.'
| name | start |
| version | 1.2.0 |
| description | Start a new session. Trigger with /start at the beginning of a conversation. Reads the session brief, checks git status, summarizes where things stand, and increments the session number. |
Run this at the beginning of every new session. It gives you and the user a shared understanding of where things left off and what's in play.
Read brief/session-brief.md in full. This is the primary source for what happened last session, what decisions were made, and what's still open.
Run git status and git log --oneline -5 to see:
If the project has an .agents/ directory, run the project-level sync from the sync-skills skill (hardlink .claude/skills/ → .agents/skills/). Do not prompt for scope — project only. Only mention it if something was added or removed.
If there is no .agents/ directory, skip silently — never create one. To opt a project in, run /sync-skills once manually and add .agents/ to its .gitignore.
Note the current date and compare it to the session brief date. If significant time has passed (more than a few days), flag it — some carryover items may be stale or already handled.
Output a concise summary for the user with these sections:
## Session [N+1] — [Current Date]
### Last Session Recap
[2-3 sentence summary of what happened]
### Open Items
[Bulleted list of carryover tasks, prioritized by urgency/relevance]
### Uncommitted Changes
[List any uncommitted files, or "Clean working tree" if none]
### Ready when you are.
Keep it tight. The user has already read the brief if they want detail — this is the quick orientation.
Note the new session number based on the brief's session number. When the user runs /brief at the end of this session, use the incremented number.
gh pr checks on it — a red CI run reframes session priorities before any planned work begins. (Session 66, 2026-07-03)