Turning Business Processes into High‑Quality Skills: A Practical Path

Business Process AI‑ization Pain Points

Users encounter scattered documentation, ambiguous workflow descriptions, loss of login state, missing local dependencies, inaccurate browser clicks, and context loss when the conversation window changes. These issues cause repeated failures when trying to automate processes with AI.

Stable Skill‑Creation Path

Use Codex Record & Replay on macOS (or record a video on Windows) to capture a manual operation and generate a SKILL.md file.

If the generated file is in English, translate it to Chinese with Codex.

Supplement the skill with business background, purpose, input sources, output targets, field definitions, exception handling, and security constraints; perform a manual review.

Replace browser actions with API calls.

Replace local application actions (e.g., Excel manipulation, file renaming, archiving) with local scripts.

Validate the skill with real data; let Codex modify the skill directly when execution fails.

Recording to Generate a Skill

Install the Record & Replay plugin, record a low‑risk read‑only operation, and ask Codex to explain each step. After any page change, re‑verify the generated description.

codex根据视频生成skill，token消耗统计：
工作内容：打开浏览器，跳转部门风险统计页面，查询并下载excel，基于Excel进行聚合统计
视频长度：1分30秒
codex推理等级：中，模型：GPT5.5
codex生成skill时长：6分钟
cc-switch统计-token消耗：$2

Upgrading Page Operations to API Calls

Export HAR and Analyze Network Requests

Open the browser DevTools Network panel, perform the workflow, and export a HAR file. The HAR contains request/response headers, cookies, and timings.

AI‑Generated Replacement Plan

Prompt Codex to read the old SKILL.md and the HAR, then output a replacement plan that identifies core business APIs, filters out static resources, logs, and outlines the suggested replacement method (API, script, manual confirmation, etc.).

请基于我提供的旧 `SKILL.md`以及 HAR 文件输出一份Skill升级替换方案。
先不要修改 Skill，也不要执行接口。请先完成分析并给出替换方案。
要求：
1. 先还原原流程…
2. 再分析网络请求…
3. 输出替换方案…
4. 登录态处理…
5. API 替换要求…
6. 本地脚本替换要求…
7. 信息不足时…
最终只输出升级方案，不要直接改文件。

Solidifying the Skill

Browser query step → call query API.

Browser export step → call export API and poll task status.

Open downloaded Excel → read with a script.

Manual copy‑paste → replace with a script for merging, filtering, and generating results.

Human glance → keep sampling verification points.

Multi‑Machine Stable Execution

Run the skill three times on the same machine: (1) with the original input, (2) with new real input, and (3) with edge‑case input. Then share the skill with a colleague or run it on another computer, checking for absolute paths, hard‑coded accounts, missing plugins, or permission issues. Finally, validate business results by sampling output against manual standards.

Timed Automatic Execution

When the skill runs stably, schedule it with Codex’s built‑in scheduler or an external task system. Include failure alerts, consider computer sleep or shutdown, and retain a final human check on key business results.

Real‑World Case: JoyHR Resume Search Skill

Original manual flow:

Open JoyHR resume page.

Select filter range.

Click search.

Iterate each resume and click export.

Download PDF.

Manual inspection.

API‑based skill flow:

Convert recruitment demand into JoyHR search parameters.

Call the talent‑pool search API, paginate candidates, retain safety summary and reviewRank.

For each candidate, call the resume‑detail API to extract work, project, education, and skill tags.

Replace per‑resume page clicks with a script that generates initial screening results and rejection reasons.

Rank candidates into “priority”, “backup”, and “reject”.

Download PDFs only for priority and backup candidates and store them in a designated folder.

Output filter criteria, candidate list, recommendation reasons, risk points, file paths, and download status.

Human confirmation.

Future Iteration

The team is progressively replacing internal web pages and software with CLI tools to further lower the cost of AI‑driven business process automation.