把一次成功变成有输入、步骤和检查点的任务卡。先验证新材料也能用, 再决定是否需要 skill 或专门 agent。Turn one success into a task card with inputs, a procedure, and checks. Test it on new material before deciding whether it needs a skill or a specialist agent.
订阅表终于算对,下个月却得重新解释退款、币种和重复行。你真正想留下的是经过确认的做法:下一次给什么材料,按什么规则处理,遇到什么情况停下来,怎样知道做完了。把这些写成一张任务卡,比保存整段长聊天更容易复用。The subscription table is finally correct, but next month you must explain refunds, currencies, and duplicates again. Preserve the method: inputs, processing rules, stopping conditions, and acceptance checks. One task card is easier to reuse than a long conversation.先从能复制粘贴的任务卡开始。反复使用而且规则稳定后,再考虑 skill。主线读者不需要编辑配置文件;附录提供希望用本地工具和多 Agent 的进阶路线。保存一种方法不会自动让它运行,也不会替你授予外部账号权限。Begin with a copyable task card. Consider a skill only after repeated use with stable rules. Main-route readers need no configuration files; the appendices cover local tools and multi-agent work. Saving a method neither schedules it nor grants external account permissions.
— I
从一次成功里抽出不变的部分Extract What Stays the Same.
回看表格练习:九月、文件名和八笔唯一交易会变;保留原始记录、分币种、记录排除项的原则不该变。把这两类内容分开,否则任务卡会把下个月的账又当作九月来算。参考答案 484.00 只能用来测试那份练习材料,不能成为未来报表的固定结论。In the spreadsheet exercise, September, the filename, and eight unique transactions are variable. Preserving originals, separating currencies, and logging exclusions are stable. Separate them so next month's data is not treated as September. The expected 484.00 belongs only to the practice fixture, never as a fixed future answer.
任务卡字段Card field
订阅核对实例Subscription review example
输入Inputs
一份流水副本、时间范围、金额正负号含义A transaction copy, period, and sign convention
处理规则Rules
分币种汇总,完全相同的同 ID 行去重;冲突待核Separate currencies; deduplicate identical same-ID rows; flag conflicts
成果Output
原始表、整理表、汇总、待核清单Raw data, cleaned data, summary, review list
停止点Stop point
字段不明或 ID 冲突先问;不连接银行、不退订Ask about unclear fields or conflicting IDs; no bank access or cancellation
验收Acceptance
行数有交代,退款计入,每个总数可回溯Account for rows and refunds; trace each total
材料来源的变化也要写进卡里。换一家平台导出,去重规则可能失效;新增一种币种,不能套用旧汇率;字段改名,先确认含义。稳定方法应当知道自己什么时候不适用,不能为了顺利执行而猜输入。Include source changes in the card. A different exporter may invalidate deduplication; a new currency cannot inherit an old exchange rate; a renamed field needs interpretation. A reusable method should identify when it does not apply instead of guessing to keep moving.
— II
先保存一张可读的任务卡Save a Card You Can Read.
提示词Prompt从已确认的工作中提炼Extract from approved work
把这次我已经确认的 <流程> 整理成一张可复用任务卡。分成输入、固定规则、每次要填的变量、输出、人工确认点和验收。只保留实际用过且认可的做法,不要把这次具体结果写成通用答案。不安装工具、不创建自动任务、不扩大权限。最后列出三个测试案例。Turn the approved <workflow> into a reusable task card: inputs, fixed rules, variables, outputs, human checkpoints, and acceptance. Keep only methods actually used and approved; do not turn this run's answer into a universal answer. Install nothing, create no schedule, and expand no permissions. Finish with three test cases.
把 subscriptions.csv 的整理规则做成「订阅流水核对 v1」任务卡。变量是输入文件、月份和正负号约定;固定规则是保留原始行、分币种、同 ID 同字段只计一次、冲突先问、退款算入现金净支出、年费折算单列、未知不猜。交付四张表,不退订。484.00 与 12.00 只放练习测试答案。给正常输入、同 ID 冲突、不相关任务三种测试,现在只写卡片。Create a Subscription review v1 card from subscriptions.csv. Variables are source file, month, and sign convention. Rules: preserve raw rows, separate currencies, count identical same-ID rows once, ask about conflicts, include refunds in net cash spending, separate annual equivalents, and preserve unknowns. Deliver four tables and cancel nothing. Put 484.00 and 12.00 only in fixture test expectations. Include normal, conflicting-ID, and unrelated-task tests. Write the card only.
把卡片保存在自己的笔记或项目资料中,标题带版本,开头写适用范围。下一次开新聊天,贴卡片和新材料即可。这个办法不依赖记忆是否开启,也便于看出它到底拿到了哪些规则。不要让账户号、真实账单或登录信息混进可分享的卡片。Save the card in your notes or project materials with a version and scope. In a new chat, supply the card and new inputs. This works without relying on enabled memory and makes the supplied rules visible. Keep account numbers, real statements, and sign-in information out of any shareable card.
— III
用三次小测试检查它会不会误用Test Correct Use and Misuse.
第一测用原练习 CSV,复核九行原始记录、八笔唯一交易、CNY 484.00 与 USD 12.00。第二测在副本把重复 T004 的金额改成 200.00,应发现同 ID 字段冲突并停下来问你。第三测只问「帮我写读书会邀请」,这张核对卡不应接管写作任务。First, use the original CSV: verify nine raw rows, eight transactions, CNY 484.00, and USD 12.00. Second, change the duplicate T004 amount to 200.00 in a copy: it should flag the same-ID conflict and ask you. Third, request a book-club invitation: this review card should not take over an unrelated writing task.这三测是读者要执行的流程验收,不是本章声称已在每种客户端跑过。若第一测答对、第二测却悄悄保留较大的金额,修订规则为「冲突时不得自行选择版本,只交冲突清单」。不要仅把例外塞在卡末尾;把它放进处理步骤和停止点,再重跑第二测。These are reader-run workflow checks, not claims of testing every client. If the normal case passes but the conflict test silently keeps the larger amount, revise the rule: never select a version automatically; return the conflict list. Put the rule in both processing and stopping conditions, then rerun that test.保存失败输入和修订前后的规则。下一次改进只解决真实遇到的问题,不预先堆十几页所有可能场景。卡片越长,维护成本越高;如果每次都必须解释一遍,先缩窄用途,例如只处理同一种导出格式。Keep the failing input and rules before and after revision. Improve real observed problems instead of anticipating every scenario across many pages. Longer cards cost more to maintain. If each run needs extensive explanation, narrow the scope to one export format.
— IV
需要经常调用,再做成 skillUse a Skill When Repetition Warrants It.
Skill 是装着工作说明及可选参考文件、脚本的包。它根据描述匹配任务,也可显式选择。当前官方建议一个 skill 聚焦一件事,先用指令,确需确定性计算再加脚本。1注 1Note 1OpenAI · Build skills。截至 2026-10-09,skill 可打包指令、资源与可选脚本,用描述匹配任务或显式调用;官方建议一项 skill 只管一件事,先用指令,并测试触发行为。OpenAI · Build skills. Checked on 2026-10-09: skills package instructions, resources, and optional scripts, with explicit or description-based invocation. Keep them focused, prefer instructions first, and test matching behavior.方法做成 skill 后,你仍然要给当次文件与授权。A skill packages instructions and optional references or scripts. It can match a task by description or be selected explicitly. Current official guidance favors one focused job and instructions first, adding scripts when deterministic work is needed.1注 1Note 1OpenAI · Build skills。截至 2026-10-09,skill 可打包指令、资源与可选脚本,用描述匹配任务或显式调用;官方建议一项 skill 只管一件事,先用指令,并测试触发行为。OpenAI · Build skills. Checked on 2026-10-09: skills package instructions, resources, and optional scripts, with explicit or description-based invocation. Keep them focused, prefer instructions first, and test matching behavior. You still supply each run's files and authorization.截至 2026-10-09,桌面端可从 Skills 查看独立技能;支持的 ChatGPT 界面用 @ 选择技能或插件,CLI / IDE 的调用语法与支持范围见附录。网页和移动端通过插件提供的技能与本机文件夹里的独立技能并不等同。先在你的界面确认可用,不要只因电脑里存了一个文件就以为手机也能调用。As of 2026-10-09, the desktop Skills view exposes standalone skills. Supported ChatGPT surfaces use @ to select skills or plugins; CLI and IDE syntax is covered in the appendices. Plugin-bundled skills on web and mobile are not the same as standalone files on your computer. Confirm availability in the actual interface before assuming a local file is usable on your phone.
提示词Prompt创建器可用时When the skill creator is available
把我提供的 <已通过测试的任务卡> 做成一个仅处理 <窄任务> 的 skill。名称要能看懂,描述写清何时用、何时不要用。默认只写指令,保留输入、输出、停止点、测试材料与验收。先展示文件清单和内容供我检查,确认后再保存到所选位置。不创建定时任务或安装外部连接。Turn <tested task card> into a skill for <narrow task> only. Use a clear name and a description stating when to use it and when not to. Start with instructions only, preserving inputs, outputs, stop points, fixtures, and checks. Show files and contents for review before saving to my chosen location. Create no schedule or external connection.
请用可用的 skill creator,把「订阅流水核对 v1」做成 subscription-review skill。仅用于我主动提供的交易表整理,不用于写邮件、投资分析、银行登录或退订。保留同 ID 冲突先问、不同币种不相加和原始数据不改。把九行合成 CSV 作为测试材料,不包含真实账单。先给我审阅,确认后保存;不要顺便配置自动运行。Use the available skill creator to turn Subscription review v1 into subscription-review. Use it only for transaction tables I supply, not email writing, investment analysis, bank sign-in, or cancellation. Preserve conflict checks, separate currencies, and immutable source data. Include the nine-row synthetic CSV as a fixture, never real statements. Show it for review before saving; do not configure automatic runs.
Description 写得太宽,容易误触发;写得太窄或缺少常用词,可能找不到。创建后新开一次聊天,显式选择该技能跑练习,再检查普通任务是否误调用。遇到失效先看它有没有实际加载规则,再查材料、工具与权限,不要用更长的提示掩盖缺失工具。A broad description can trigger on unrelated tasks; a narrow or obscure one may never match. Test explicit selection in a new chat and check unrelated tasks. If it fails, first establish whether the instructions loaded, then inspect inputs, tools, and permissions. A longer prompt cannot substitute for a missing tool.
— V
录一次操作,也要检查录下了什么Review What a Recording Captures.
支持的 macOS 账号在启用 Computer Use 后,可用 Record & Replay 演示流程并生成 skill;截至 2026-10-09 不应写成跨平台通用功能。2注 2Note 2OpenAI · Record & Replay。截至 2026-10-09,该录制方式要求 macOS 且 Computer Use 可用并启用;录制完成后生成可继续修订的 skill,不等于每次重放都必然成功。OpenAI · Record & Replay. As of 2026-10-09, this requires macOS and available, enabled Computer Use. A recorded workflow becomes a revisable skill, not a guarantee of successful replay.它适合稳定、短而完整的操作,例如把已脱敏的报告下载到指定练习目录。On supported macOS accounts with Computer Use enabled, Record & Replay can turn a demonstrated workflow into a skill. As of 2026-10-09, it is not a cross-platform promise.2注 2Note 2OpenAI · Record & Replay。截至 2026-10-09,该录制方式要求 macOS 且 Computer Use 可用并启用;录制完成后生成可继续修订的 skill,不等于每次重放都必然成功。OpenAI · Record & Replay. As of 2026-10-09, this requires macOS and available, enabled Computer Use. A recorded workflow becomes a revisable skill, not a guarantee of successful replay. Use a short, stable, complete task, such as downloading a redacted report into a practice folder.录制前关闭无关窗口,避开密码和私人聊天,说清日期、文件名等每次会变的值。完成立即停止录制,检查生成说明有没有把你的临时点击误当成规则。界面改版后旧坐标和旧菜单可能失效,重放仍要看结果。没有录制入口,手工任务卡已经足够,不必为了录一次屏升级账号。Close unrelated windows, avoid passwords and private conversations, and identify variables such as dates and filenames before recording. Stop as soon as the task ends. Review whether accidental clicks became instructions. Interface changes can invalidate old steps, so inspect replay results. Without recording access, a manual task card is enough.
— VI
方法、分工、频率各管一件事Keep Method, Delegation, and Timing Separate.
一个九行表格不需要三个 agent。材料真的很多时,可以让两个只读助手分别查来源与数字,各回短摘要和证据位置,由主任务统一改稿。不要让多个助手同时改同一份成品。完整多 Agent 配置与账见进阶附录,这里先判断分工有没有减少等待,有没有增加核对和用量。A nine-row table does not need three agents. With genuinely large material, two read-only helpers can check sources and numbers separately, returning concise findings and evidence locations while the main task owns edits. Avoid concurrent edits to one deliverable. The multi-agent appendix covers configuration and cost; first ask whether delegation reduces waiting enough to justify extra checking and usage.保存好方法后,若真有固定时间要重复,再去下一章设 schedule。3注 3Note 3OpenAI · Scheduled tasks。保存工作方法与创建定时任务是不同步骤;应先在普通聊天手动验证,再配置频率并检查前几次运行。核对于 2026-10-09。OpenAI · Scheduled tasks. Saving a method and creating a schedule are separate steps. Test in a regular chat, then schedule and review the first runs. Checked on 2026-10-09.每月手动做一次也完全合理。先连续使用两次并检查结果,记下卡片版本、输入格式和最后验证日期。规则过时或长期不再用,停用或归档,让可复用清单保持短而可信。If the saved method really needs recurring execution, set a schedule in the next chapter.3注 3Note 3OpenAI · Scheduled tasks。保存工作方法与创建定时任务是不同步骤;应先在普通聊天手动验证,再配置频率并检查前几次运行。核对于 2026-10-09。OpenAI · Scheduled tasks. Saving a method and creating a schedule are separate steps. Test in a regular chat, then schedule and review the first runs. Checked on 2026-10-09. Running it manually once a month is also reasonable. Use and inspect it twice, recording the card version, input format, and last verification date. Disable or archive stale methods so the reusable collection stays small and trustworthy.
留下经过检查的做法,
也留下它该停下来的地方.
Keep the method that worked,
and the conditions that make it stop.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return
