Capture a Manual Task as a Reusable Skill

An unverified recipe: demonstrate once, capture a reusable skill card, then check it with a dry-run.

Who

People who repeat backend or browser-based operations

You get

A reusable skill card with placeholder-driven steps and guardrails

You need

Grok Bot

Steps

1. Demonstrate once

   Perform the task fully while the bot observes or reads your supplied log. Do not let the bot jump in mid-stream. Narrate goals and success criteria aloud so intent is captured, not just clicks.

2. Extract the skill card

   The bot outputs: objective, prerequisites, numbered verifiable steps, success criteria, failure/rollback language, and a "forbidden actions" block. Every variable becomes a {{placeholder}} with a description and example fake value. No invented clicks or commands beyond what appeared in the demonstration.

   Execution prompt · Source text。Grok Bot

   Copy prompt

   你是技能整理员。根据以下「一次完整示范」记录，生成可复用技能卡。
   
   任务名称：{{任务名称}}
   
   示范记录：
   {{示范步骤或文字稿}}
   
   输出结构（Markdown）：
   # {{任务名称}}
   ## 目标（一句话）
   ## 前置条件
   ## 输入占位符（表格：名 / 含义 / 示例假值）
   ## 步骤（编号，每步可验证）
   ## 成功标准
   ## 失败与回滚
   ## 禁止事项（必须包含：不写 key；不对外发送；不删未确认文件）
   
   规则：
   - 真实绝对路径、邮箱、密钥改成 {{占位符}}
   - 不要发明示范里没出现的点击或命令
   - 有副作用的步骤标 [需人工确认]
   - 若示范含糊，列出「待澄清」而不是猜测

3. Sanitize and tighten

   You edit one round: strip internal absolute paths, real domains, and account names. Add explicit "do not" rules. Flag any vague step as "needs clarification" rather than letting the bot guess.

4. Dry-run with fake data

   Use fictional inputs to produce an execution plan without performing real actions. Supply the skill card and fictional variables with the original prompt below. Without confirmation, the prompt permits only a dry-run; compare the plan with your intended result before a live pilot.

   Execution prompt · Source text。Grok Bot

   Copy prompt

   按技能卡 {{技能卡路径或粘贴}} 执行任务。
   本次变量：
   {{占位符赋值表}}
   
   先输出「计划」：将执行的步骤摘要 + 副作用列表。
   等待人回复「确认」后再执行；未确认只准干跑。
   执行中每步写简短日志；偏离技能卡时停止并询问。

5. Live pilot (optional)

   Optional: choose one low-risk instance and reuse the complete execution prompt from the previous step. Review its plan before explicitly confirming. Log each step, stop and ask on deviations, and write any correction back into the skill card rather than leaving it only in the conversation.

6. Store and version

   Deposit the sanitized skill card in a private skills directory or, if generic enough, in an examples/ folder of a public repo. Public versions must contain no org-specific data. Tag with creation date and last-verified date. Fictional output example supplied by the source, preserved in Chinese; this is not an execution result: ```markdown # 发布前检查 · AcmeTools 文档站 ## 目标 在宣布更新前，确认公开文档链接可打开且无占位符残留。 ## 输入占位符 | 名 | 含义 | 示例假值 | | {{文档根URL}} | 公开文档站点 | https://docs.example.com | | {{版本号}} | 本次版本标签 | v0.0.0-demo | ## 步骤 1. 打开 {{文档根URL}}/changelog 确认 {{版本号}} 条目存在 2. 抽查 3 个内链可打开 3. 全文搜索 TODO / FIXME / {{ 是否残留 4. 输出检查表给人对勾；[需人工确认] 后方可宣布 ## 禁止事项 不修改生产配置；不发送公告；不提交 git push ```

7. Establish calling convention

   For future runs: load the skill card first, then execute. If the conversation contradicts the card, the card plus your live override wins—but update the card afterward so the source of truth stays current.

你是技能整理员。根据以下「一次完整示范」记录，生成可复用技能卡。

任务名称：{{任务名称}}

示范记录：
{{示范步骤或文字稿}}

输出结构（Markdown）：
# {{任务名称}}
## 目标（一句话）
## 前置条件
## 输入占位符（表格：名 / 含义 / 示例假值）
## 步骤（编号，每步可验证）
## 成功标准
## 失败与回滚
## 禁止事项（必须包含：不写 key；不对外发送；不删未确认文件）

规则：
- 真实绝对路径、邮箱、密钥改成 {{占位符}}
- 不要发明示范里没出现的点击或命令
- 有副作用的步骤标 [需人工确认]
- 若示范含糊，列出「待澄清」而不是猜测