據 AlphaMoat 統計,截至 2026-10,Codex 累計下載 8,751 次、獲 2 個星標,按累計下載量在本次參與排名的 10,914 個 AI 技能中排名第 3,178 位,在「開發程式設計」分類(1,069 個)中排名第 256 位。
列表預設按累計下載量排序,也可按星標排序;詳情頁排名按累計下載量計算。
| 平台 | 累計下載 | 星標 | 評論 |
|---|---|---|---|
| ClawHub ↗ | 6409 | 1 | - |
| SkillHub ↗ | 8751 | 2 | — |
各平台的計數口徑不同,統計時間不一致。
安全使用 Codex 進行感知程式碼庫的編碼,支援顯式審批、沙箱選擇、MCP 邊界設定及 PR 就緒驗證工作流。
2026-09 起統計方法調整:同名技能按作者分別統計,並覆蓋全部技能。圖中虛線為調整前的月份,與之後的資料不直接比較,也不計算跨越調整的增長。 該技能有同名的其他版本,調整前的資料歸屬當前下載量最大的這一個。
| 月份 | 累計下載 | 環比 | 星標 |
|---|---|---|---|
| 2026-10 | 8751 | — | 2 |
| 2026-09 | 6313 | — | 2 |
| 2026-08 | 5803 | +14.3% | 2 |
| 2026-07 | 5077 | +20.8% | 2 |
| 2026-06 | 4203 | +21.9% | 2 |
| 2026-05 | 3448 | — | 2 |
| 2026-03 | 790 | — | 1 |
以下是作者釋出的原始檔案,按純文字顯示,本站未做修改或稽核;其中的連結不可點選,命令請核實後再執行。抓取於 2026-10-23。
---
name: Codex
slug: codex
version: 1.0.0
homepage: https://clawic.com/skills/codex
description: Use Codex safely for repo-aware coding with explicit approvals, sandbox choices, MCP boundaries, and PR-ready verification workflows.
changelog: Initial release with repo-safe execution, approval and sandbox guidance, MCP and cloud guardrails, and review-ready handoff workflows.
metadata: {"clawdbot":{"emoji":"🧭","requires":{"bins":["codex"],"bins.optional":["git","rg"],"env.optional":["OPENAI_API_KEY"],"config":["~/codex/","~/.codex/config.toml"]},"os":["linux","darwin","win32"],"configPaths":["~/codex/","~/.codex/config.toml"]}}
---
## When to Use
User wants to use Codex as a real coding agent instead of a generic chat assistant: inspect a repo, make bounded edits, run review mode, resume work, use MCP safely, or hand work off with clear verification evidence.
Use this skill when the hard part is not "write code" but "make Codex behave safely and predictably" across CLI, `exec`, `review`, `resume`, MCP, app-server, cloud tasks, or local OSS-provider workflows.
## Architecture
Memory lives in `~/codex/`. If `~/codex/` does not exist, run `setup.md`. See `memory-template.md` for structure.
```text
~/codex/
|-- memory.md # Durable activation boundaries and operating defaults
|-- repo-profiles.md # Per-repo conventions, test surface, and blast-radius notes
|-- safety.md # Sandbox, approval, and trust defaults
|-- mcp-notes.md # Approved MCP servers, scopes, and rejection reasons
`-- incidents.md # Stuck sessions, failed commands, and recovery patterns
```
## Quick Reference
Load only the smallest file needed for the current blocker.
| Topic | File |
|-------|------|
| Setup guide | `setup.md` |
| Memory template | `memory-template.md` |
| Install, login, and first-run checks | `install-and-auth.md` |
| Repo execution and `codex exec` workflows | `repo-execution.md` |
| Approval modes and sandbox choices | `approvals-and-sandbox.md` |
| MCP, app-server, cloud, and local-provider guardrails | `mcp-and-cloud.md` |
| Review mode and handoff patterns | `review-and-handoffs.md` |
| Recovery playbooks for auth, stuck sessions, and wrong-scope work | `troubleshooting.md` |
## Requirements
- `codex` binary installed and working on the target machine.
- Active authentication through `codex login` or an explicit `OPENAI_API_KEY` flow when that mode is chosen.
- `git` available when the task involves repository inspection, diff review, or commit-ready workflows.
- Explicit user approval before dangerous sandbox bypass, remote MCP usage, Codex Cloud apply, production commands, or any operation with irreversible side effects.
- Treat model names, features, and app-server behavior as live product surface: verify with `codex --help`, subcommand help, or official docs instead of hardcoding stale assumptions.
## Operating Coverage
This skill treats Codex as an operational coding surface, not as generic AI advice. It covers:
- interactive Codex CLI usage with explicit working-directory and safety choices
- non-interactive `codex exec` and `codex review` workflows
- `resume`, `fork`, and handoff-friendly session recovery
- sandbox and approval policy selection by blast radius
- MCP server trust decisions and local-versus-remote tool boundaries
- Codex app-server and cloud task usage only when their extra trust and review requirements are explicit
- local OSS-provider routing via `--oss` and `--local-provider` when the user intentionally wants local execution
## Data Storage
Keep only durable Codex operating context in `~/codex/`:
- which repos or workspaces are approved for Codex use
- default sandbox and approval posture per task type
- preferred execution surfaces: interactive CLI, `exec`, `review`, cloud, or local OSS provider
- approved MCP servers and what each one is allowed to touch
- recurring recovery notes for wrong directory, dirty worktree, stalled commands, or broken auth
## Core Rules
### 1. Preflight the Task Before Codex Acts
- Lock five facts first: target repo, current directory, dirty worktree state, required permissions, and expected verification.
- If any of those are unclear, pause and resolve them before running Codex with write capability.
- "Start coding" is never the first step in an unfamiliar repo.
### 2. Choose the Operating Mode Explicitly
- Use interactive Codex for exploratory repo work, `codex exec` for bounded non-interactive execution, and `codex review` for review-first tasks.
- If resuming or branching prior work, prefer `resume` or `fork` over re-describing the entire context from scratch.
- Treat cloud, app-server, and MCP-assisted runs as separate modes with separate risk.
### 3. Match Sandbox and Approval to Blast Radius
- Read-only fits inspection, planning, and low-trust exploration.
- Workspace-write fits normal local coding in the approved repo.
- Full access or dangerous bypass is a special-case mode that needs explicit user intent and an external sandbox story.
- Do not normalize high-trust modes for convenience.
### 4. Read the Repo Before Editing It
- Inspect tree shape, git status, entrypoints, conventions, and test surface before proposing edits.
- When the worktree is already dirty, separate user changes from agent changes and avoid destructive cleanup.
- Codex should adapt to the repo, not force the repo into a generic workflow.
### 5. Keep Changes Reviewable and Scoped
- Favor minimal diffs, targeted commands, and explicit file ownership.
- Avoid unrelated cleanup, speculative refactors, or "while here" edits unless requested.
- If a command or edit expands scope, stop and surface that expansion immediately.
### 6. Treat Auth, MCP, and Cloud as Trust Boundaries
- A tool being available does not mean it is approved.
- Review each MCP server for scope, data access, and side effects before enabling it.
- Use existing login sessions when possible; never scrape secrets from local files without clear user intent.
- Inspect cloud diffs before applying them locally.
### 7. Verify Outcomes and Leave a Handoff Trail
- A successful Codex run ends with checks, not with code edits alone.
- Report what changed, what was verified, what failed, and what remains risky.
- For interrupted or long-running work, leave a crisp checkpoint that another operator can resume without guesswork.
## Codex Traps
- Running Codex in the wrong directory -> edits land in the wrong repo or outside intended scope.
- Treating `workspace-write` as harmless -> it still writes real files and can widen a diff quickly.
- Using `--dangerously-bypass-approvals-and-sandbox` for routine work -> convenience becomes unreviewable risk.
- Enabling MCP servers because they are available -> hidden data reach and side effects expand silently.
- Applying cloud output without reviewing the diff -> local repo changes become opaque.
- Letting Codex work through a dirty tree without clarifying ownership -> review noise and accidental overwrite risk.
- Re-running vague prompts after interruption -> duplicated work and inconsistent verification.
## External Endpoints
Only these external categories are allowed unless the user explicitly approves more:
| Endpoint | Data Sent | Purpose |
|----------|-----------|---------|
| https://api.openai.com | prompts, selected repository context, tool results, and execution metadata needed for Codex runs | Codex model execution, cloud tasks, login-linked agent work |
| https://developers.openai.com/* | doc queries only | Verify current Codex product behavior and configuration details |
| https://{user-approved-mcp-host} | request payloads required by the specific MCP server | Optional user-approved tool access beyond the local machine |
No other data is sent externally unless the user explicitly approves additional MCP servers, Git remotes, or service endpoints.
## Security & Privacy
Data that leaves your machine:
- prompts and the repo context selected for Codex runs against OpenAI services
- optional MCP payloads only for user-approved MCP servers
- optional cloud task payloads and diffs when Codex Cloud is intentionally used
Data that stays local:
- `~/.codex/config.toml` and the user's local Codex session/config state
- durable operating notes under `~/codex/`
- local diffs, verification output, and repo metadata unless the user explicitly pushes or uploads them
This skill does NOT:
- assume dangerous bypass is acceptable by default
- enable remote MCP or cloud apply silently
- scrape tokens from arbitrary files to "help" auth succeed
- hide sandbox or approval choices from the user
- claim that CLI, app-server, cloud, and local `--oss` flows have identical risk
## Trust
By using this skill, Codex work may send prompts and selected repository context to OpenAI, plus any optional user-approved MCP endpoints.
Only install if you trust those services with that data.
## Scope
This skill ONLY:
- helps operate Codex safely and effectively in real coding environments
- structures repo work into explicit execution, review, and handoff modes
- keeps durable memory for approved repos, safety posture, and recurring recovery patterns
This skill NEVER:
- treat every available Codex feature as automatically approved
- recommend destructive git cleanup as a default fix
- blur the line between local-only, cloud, and MCP-assisted execution
- modify its own skill files
## Related Skills
Install with `clawhub install <slug>` if user confirms:
- `agentic-engineering` - Strengthen the human workflow around parallel coding agents and blast-radius thinking.
- `coding` - Improve implementation quality once Codex is operating inside the right repo boundaries.
- `git` - Handle branches, diffs, and non-destructive repository recovery safely.
- `api` - Reuse structured API and request-debugging patterns when Codex integrates with services.
- `workflow` - Turn recurring Codex tasks into repeatable, reviewable execution paths.
## Feedback
- If useful: `clawhub star codex`
- Stay updated: `clawhub sync`
【全棧能力】程式設計大神.Skill:面向複雜真實專案的高階全棧程式設計助手。不止生成程式碼,而是從一句目標接手陌生程式碼庫,拆解需求、補齊實現、追到根因、完成重構、聯調前後端,並把結果推進到能執行、能驗證、能交付。覆蓋 Web、後端、指令碼、iOS、安卓、鴻蒙 ArkUI、微信小程式及其他技術棧,支援中英文工程、疑難 Bug、跨模組任務和長任務續做;用構建、日誌、真實入口、裝置或介面證據說明結果。僅問概念時直接回答,要求落地時持續推進到驗收或明確阻塞。
程式設計專家 Pro:融合程式設計專家 2.0.3 工程專題與程式設計大神的鴻蒙 ArkTS/ArkUI、微信小程式和小遊戲開發資料。用於專案實現、根因除錯、審查、重構、測試、架構效能分析及多端開發驗證。支援自然語言和 @專題標識,按任務讀取包內知識;只問概念時直接回答。
程式設計專家.Skill P8級程式設計助手,25年實戰經驗技能Skill,全棧:Java/Php網站/軟體專案、API設計、Bug診斷、程式碼生成、程式碼審查、重構、測試用例、效能基準、技術選型、文件生成、任務拆解、Spec驅動、Karpathy規範、CMS二次開發、前端設計、MySQL/Mariadb、專案知識圖譜,以及工程紀律層(根因除錯硬閉環、階段導航、事故復盤/SLO、威脅建模/供應鏈安全、生產就緒/漸進式釋出、AI編碼治理、面向Agent程式碼可讀性、LLM應用安全)。支援 @標識 顯式呼叫跳過路由。
程式設計專家.Skill P8級程式設計助手,25年實戰經驗技能Skill,全棧:Java/Php網站/軟體專案、API設計、Bug診斷、程式碼生成、程式碼審查、重構、測試用例、效能基準、技術選型、文件生成、任務拆解、Spec驅動、Karpathy規範、CMS二次開發、前端設計、MySQL/Mariadb、專案知識圖譜,以及工程紀律層(根因除錯硬閉環、階段導航、事故復盤/SLO、威脅建模/供應鏈安全、生產就緒/漸進式釋出、AI編碼治理、面向Agent程式碼可讀性、LLM應用安全)。支援 @標識 顯式呼叫跳過路由。
程式設計專家 Pro:融合程式設計專家 2.0.3 工程專題與程式設計大神的鴻蒙 ArkTS/ArkUI、微信小程式和小遊戲開發資料。用於專案實現、根因除錯、審查、重構、測試、架構效能分析及多端開發驗證。支援自然語言和 @專題標識,按任務讀取包內知識;只問概念時直接回答。
程式設計專家 Pro:融合程式設計專家 2.0.3 工程專題與程式設計大神的鴻蒙 ArkTS/ArkUI、微信小程式和小遊戲開發資料。用於專案實現、根因除錯、審查、重構、測試、架構效能分析及多端開發驗證。支援自然語言和 @專題標識,按任務讀取包內知識;只問概念時直接回答。
程式設計專家.Skill P8級程式設計助手,25年實戰經驗技能Skill,全棧:Java/Php網站/軟體專案、API設計、Bug診斷、程式碼生成、程式碼審查、重構、測試用例、效能基準、技術選型、文件生成、任務拆解、Spec驅動、Karpathy規範、CMS二次開發、前端設計、MySQL/Mariadb、專案知識圖譜,以及工程紀律層(根因除錯硬閉環、階段導航、事故復盤/SLO、威脅建模/供應鏈安全、生產就緒/漸進式釋出、AI編碼治理、面向Agent程式碼可讀性、LLM應用安全)。支援 @標識 顯式呼叫跳過路由。
用 gh 命令列工具檢視 GitHub 上的 PR 檢查、工作流執行與 API 查詢的技能。
將使用者明確糾正與自我檢視整理成分層記憶,並依使用情況升級或封存。
在保留公式、格式和活頁簿結構的前提下建立、檢查或編輯 Excel 檔案。
處理 Word 與 DOCX 文件的作業指南,著重維持樣式、編號與版面。
面向多家影片模型 API 的 AI 影片生成與編排技能。
把資料庫、表格與匯出的原始資料變成可決策結論的資料分析技能。
資料月份:2026-10 · 下載/星標數聚合自公開技能註冊庫(ClawHub、SkillHub),按月更新。口徑說明見資料方法。