把工作流教给 Genie One:技能终极指南

像 Genie 接手新任务的第一周那样测试它:按照人们真实的提问方式提问,对照黄金标准检查结果,并在对话中纠正它。
每个团队都有支配其工作方式的“配方”。财务如何汇总月末报告包,销售运营在预测电话会议前会做哪些检查,领导层在复盘时期望看到什么格式。这些模板存在于人们的头脑里、旧邮件中,以及没人找得到的“我是这么做的”文档里。
Genie One 已经对你的业务了解很多。通过 Genie Ontology,它不仅学习你的业务定义,还了解你的团队如何使用他们的数据。Skills 增加了一层额外的控制:你希望事情如何完成。一个技能是一个工作流配方,Genie One 可以按需遵循,确保每次格式和输出一致。例如,借助技能,用户可以提示 Genie One“整理我们上个月交易报告”,并知道它会以你的团队会准备该报告的方式返回,包含同样的检查、同样的结构和同样的详细程度。
Genie One 会在技能被用户直接调用时,或当它判断技能与你请求的内容相关时,运用这些技能。这意味着你可以持续添加技能,每个技能都可以包含其任务所需的尽可能多的细节,而它们的指导会限定在该任务内,不会溢出到你的其他工作中。
技能是持久、可复用的资产。一个技能就是以平实语言、采用开放的 Agent Skills 格式,对你团队工作方式所做的书面记录,因此它归你所有,可供你保留、审阅并在此基础上继续构建。通过 Unity Gateway 把它发布出去,同一套做法就能被任何支持 MCP 的智能体使用,包括你团队已经在用的编程智能体。而一旦它跑通了,就把它分享到整个组织,让一个人的最佳工作方式成为所有人的。
上手技能很简单。你只需在 Genie One 中通过一次对话,把工作流程详细解释一遍,就能直接构建出一个技能。
何时该动用一项技能
当你发现自己不止一次向 Genie One 解释同一件事时,就该考虑做一个技能了。以下是一些好迹象:
它定期发生:每周更新、针对单个客户的调查、按请求生成的工单、会前摘要。
它有步骤:拉取这个,与那个对比,然后写成文档。
格式很重要:某个特定数字必须始终包含在内,管理层期望先看到标题,其次是表格,最后是风险。
有些规则你总会用到:什么算作异常、哪些数字要四舍五入、如何签字确认。
你一直在手动做同样的调整:如果每个响应都需要同样的调整,那么这个调整步骤就应该放进 Skill 里。
Skills 与 Genie One 了解你组织的其他方式并列。
你的业务定义应放在 Genie Ontology 中。你在 Unity Catalog Semantics 中整理最重要的定义,而 Genie Ontology 会将它们与它从你的数据已被如何使用中学到的上下文结合起来。
Metric views 保存可复用的结构化指标,例如“净销售额”或“毛利率”是如何计算的。
Pages 保存概念和分类体系的书面定义,例如什么算作“活跃客户”或“合格销售线索”。
Domains 按业务领域对两者进行组织,例如财务或销售。
适用于每次对话的约定应放在工作区指令中。
个人事实,例如你的角色或报告货币,应放在记忆中。
技能承载的是方法:一项工作的步骤、检查项和格式。Genie Ontology 告诉 Genie One 各种事物意味着什么,而技能则告诉它以某种特定方式把某件事做成。
技能还与 Genie Agents 协同工作。Genie Agent 是你组织中某个领域的专家,它掌握某个限定范围内的指标定义、数据表、关联关系和业务规则,比如销售、财务或人力资源。技能则是某项具体任务的做法:步骤、检查清单和输出格式。举例来说,销售管道的 Genie Agent 了解管道相关的数据表、关联关系和阶段定义,而“每周预测风险备忘录”这项技能则告诉 Genie One 将本周管道与上周对比、标出发生滑单的交易,并生成一页备忘录。Genie Agent 提供领域知识,技能则为 Genie One 提供工作流程。
先梳理工作流,再动手构建
技能的质量取决于你对这项工作的描述有多到位。开始之前,先把任务从头到尾走一遍,并回答几个问题:
它应该在什么条件下被使用?当有人说什么或问什么时,就意味着这项技能适用?
它使用哪些工具和数据?涉及哪些数据、Genie Agents、文档,或 Gmail、Google Drive 等已连接的工具?
各个步骤按什么顺序发生?有些步骤依赖其他步骤,比如先确认哪个月份已经完整,再去拉取该月的数据。
哪些地方需要深入挖掘?有些步骤应反复执行,直到某个问题得到解答,比如把一个异常结果层层拆解,直到能用一句话解释清楚。
每次都要检查什么?检查覆盖的不只是数字。它们也可以是结果必须满足的条件,比如文档始终包含你团队期望的那四列表格。
结果应该长什么样?输出是一份书面摘要?一封邮件?一张需要归档的工单?包括分节、长度、语气,以及最终落到哪里。
你希望在哪些环节有发言权?决定这项技能是直接给出最终结果,还是暂停等你审核,或者先回答一些前置问题。
然后把它交给 Genie One。用你自己的话描述工作流,并要求它在动手编写之前先采访你。Genie One 借助 Genie Ontology 调用它已经掌握的信息,因此它的问题聚焦于你的工作方式,而不是你的数据意味着什么。
示例:构建月度经营简报
一个商业财务团队每个月为管理层撰写一份经营简报。这个工作流的描述很简单:拉取上个月的数字,按几条规则逐一检查每个渠道,深入排查任何异常之处,查找团队中解释这些异常的邮件,然后写一份简短的简报。
仅用几句话描述,就足以让 Genie One 开始工作。它自行通过 Genie Ontology 找到了相关的 Genie Agent 和受治理的指标视图,随后询问具体方法:每个月哪些指标重要、要展示哪些拆分维度、什么算异常、书面内容应该是什么样,以及是否要暂停等待审核。用平实的语言回答,就足以让 Genie One 写出这项技能。
决定每个环节应该有多严格
判断力,才是让一项技能超越定时报告的关键。仪表盘或固定作业每次做的都是同样的事。Genie One 可以借助技能来做出判断,比如挑出最值得管理层关注的三个变化,对异常结果一路追查直到弄清原委,并用通俗的语言解释清楚——就像你会放心地把这些交给一位经验丰富的同事去做一样。
但这并不意味着一切都该敞开不管。有些部分必须每次输出都完全一致,另一些则更适合交给 Genie One 去判断。要认真想清楚哪些该归哪一类。凡是别人要依赖的内容,都适合用严格规则固定下来:需要跨时间比较的数字、其他团队或系统所预期的格式、决定哪些内容需要进一步细查的阈值,以及必须始终存在的免责声明或签核。
条件也值得明确写出来:哪些内容必须始终出现,以及当某些内容缺失时应当如何处理。例如,交易简报始终包含观察清单,即便它只是写着没有任何指标突破阈值,因为缺失的章节看起来就像是漏掉了某个步骤。
小型技能,串联起来
由于 Genie One 只读取自己需要的技能,把技能做得具体而聚焦往往更有利。一项只把一件事做好的技能,比一项试图包揽一切的技能更容易测试、更容易修复,也更容易复用。真正的收益在于,小型技能可以串联起来。
一个有用的思考方式:
有些技能负责完成一项任务,比如生成交易简报。
有些技能承载共享的专业知识,比如你们团队面向管理层文档的风格规范。
有些技能会按顺序调用其他技能,比如一个月末打包技能,先运行交易简报,再运行现金汇总,然后把两者合并在一起。
举例来说,只要更新一次风格指南技能,所有使用它的其他技能都会自动带上这一改动。先从一个技能开始,当指令开始覆盖两项不同的工作时,或者当你发现同样的规则被复制进第二个技能时,再把它拆分。
像测试 Genie 第一周接手新任务那样去测试它
给 Genie One 加一项新技能,就像迎来一位能干的新员工:聪明、上手快,但你还是要检查他最初做的几件事,确认交付的正是你想要的东西。要按人们真正使用它的方式来测试这项技能。用你自己的话去问,而不是直接叫它的名字,并对照你所了解的情况核查结果。
几个习惯能让测试更快:
原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Teach Genie One your workflows: The ultimate guide for skills in Genie One)
Test it like Genie's first week on a new task: Ask the way people really ask, check the result against a golden reference, and correct it in conversation.
Every team has “recipes” that govern how they work. The way finance puts together the month-end pack, the checks sales operations runs before a forecast call, the format leadership expects when performing a retrospective. These templates live in people's heads, in old emails and in "here's how I do it" documents that nobody can find.
Genie One already understands a lot about your business. Through Genie Ontology, it not only learns your business definitions, but also how your teams use their data. Skills add an extra layer of control: how you like things done. A skill is a workflow recipe that Genie One can follow on demand ensuring a consistent format and output each time. For example, with a skill, users can prompt Genie One to "Put together our trading report for last month", and know that it will return the report the way your team would have prepared it, with the same checks, the same structure and the same level of detail.
Genie One leverages skills when they are either called directly by the user, or when it decides that it's relevant to what you've requested. That means you can keep adding skills, each with as much detail as its task needs, and their guidance stays with that task rather than spilling into your other work.
Skills are durable, reusable assets. A skill is a written record of how your team works, in plain language and in the open Agent Skills format, so it's yours to keep, review and build upon. Publish it through Unity Gateway and the same recipe can be used by any agent that supports MCP, including the coding agents your teams already use. And once it works, share it across your organisation, so one person's best way of working becomes everyone's.
It’s easy to get started with skills. You can build a skill directly in Genie One through one conversation, simply by explaining your workflow in detail once.

When to reach for a skill
You should consider a skill when you find yourself explaining the same thing to Genie One more than once. Good signs include:
It’s done regularly: a weekly update, a per-client investigation, a per-request ticket, a pre-meeting summary.
It has steps: pull this, compare it with that, then write it up.
The format matters: a specific figure always needs to be included, leadership expects a headline first, a table second and risks last.
There are rules you always apply: what counts as unusual, which numbers to round, how you sign off.
You keep manually making the same tweak: if every response needs the same tweak, that adjustment step belongs in a skill.
Skills sit alongside the other ways Genie One learns your organisation.
Your business definitions belong in Genie Ontology. You curate the most important ones in Unity Catalog Semantics, and Genie Ontology combines them with context it learns from how your data is already used.
Metric views hold reusable, structured metrics, such as how "net sales" or "gross margin" is calculated.
Pages hold written definitions of concepts and taxonomy, such as what counts as an "active customer" or a "qualified sales lead".
Domains organise both by business area, such as finance or sales.
Conventions that apply to every conversation belong in workspace instructions.
Personal facts, like your role or your reporting currency, belong in memory.
A skill holds the method: the steps, checks and format for a piece of work. Genie Ontology tells Genie One what things mean, and a skill tells it a specific way to get something done.
Skills also work hand in hand with Genie Agents. A Genie Agent is the domain expert for one area of your organisation. It knows the metric definitions, tables, joins and business rules for one scoped domain, such as sales, finance or HR. A skill is the recipe for a specific task: the steps, the checklist and the output format. For example, a sales pipeline Genie Agent knows the pipeline tables, joins and stage definitions. A "weekly forecast risk memo" skill tells Genie One to compare this week's pipeline with last week's, flag deals that slipped and produce a one-page memo. The Genie Agent supplies the domain knowledge and the skill supplies Genie One with the workflow.
Map the workflow before you build
The quality of a skill comes from how well you describe the work. Before you start, walk through the task end to end and answer a few questions:
Under what conditions should it be used? What would someone say or ask that means this skill applies?
What tools and data does it use? Which data, Genie Agents, documents or connected tools such as Gmail or Google Drive are involved?
What order do things happen in? Some steps depend on others, such as confirming which month is complete before pulling its numbers.
Where does it need to dig deeper? Some steps should repeat until a question is answered, like breaking an unusual result down until it can be explained in one sentence.
What gets checked every time? Checks cover more than the numbers. They can also be conditions the result has to meet, such as the document always containing a table with the four columns your team expects.
What should the result look like? Is the output a written summary? An email? A ticket that needs to be filed? The sections, the length, the tone and where it ends up.
Where do you want a say? Decide whether the skill should go straight to a finished result or pause for you to review, or answer some upfront questions.
Then hand it to Genie One. Describe the workflow in your own words and ask it to interview you before it writes anything. Genie One draws on what it already knows through Genie Ontology, so its questions focus on how you work rather than what your data means.
Example: building a monthly trading brief
A commercial finance team writes a trading brief for leadership every month. The workflow is simple to describe: pull last month's numbers, check each channel against a few rules, dig into anything unusual, look for emails from the team that explain it, then write a short brief.
Described in a couple of sentences, that was enough for Genie One to get started. It found the relevant Genie Agent and governed metric view through Genie Ontology on its own, then asked about the method: which measures matter every month, which breakdowns to show, what counts as unusual, what the write-up should look like and whether to pause for review. Answering in plain language was enough for Genie One to write the skill.
Decide how strict each part should be
Judgement is what makes a skill more than a scheduled report. A dashboard or a fixed job does the same thing every time. Genie One can use a skill to make decisions, like choosing the three changes most worth leadership's attention, following an unusual result until it makes sense and explaining it in plain language, the same way you would trust an experienced colleague to.
That doesn’t mean everything should be left open. Some parts need to come out exactly the same every time, whereas others are better left to Genie One's judgement. Be deliberate about which is which. Strict rules suit anything someone else relies on: numbers that are compared over time, a format another team or system expects, the thresholds that decide what gets a closer look and the disclaimers or sign-offs that must always be there.
Conditions are worth spelling out too: what must always appear, and what should happen when something is missing. For example, the trading brief always includes a watch list, even when it only says nothing breached a threshold, because a missing section looks like a missed step.
Small skills, chained together
Because Genie One only reads the skills it needs, it can be advantageous to keep them specific and focused. A skill that does one job well is easier to test, easier to fix and easier to reuse than one that tries to do everything. The real payoff is that small skills can be chained together.
A useful way to think about it:
Some skills do a task, like producing the trading brief.
Some skills hold shared know-how, like your team's style guidelines for leadership documents.
Some skills run other skills in order, like a month-end pack that runs the trading brief, then a cash summary, then puts them together.
For example, from updating the style guidelines skill once, every other skill that uses it picks up the change. Start with one skill, then split it when the instructions start covering two different jobs, or when you notice the same rules being copied into a second skill.
Test it like Genie's first week on a new task
Genie One with a new skill is like a capable new starter: smart and quick to learn, but you check their first few pieces of work to make sure they're delivering what you're looking for. Test the skill the way people will really use it. Ask in your own words rather than by name, and check the result against what you know.
A few habits make testing quicker:
Keep a golden reference: once a result comes back the way you want it, ask Genie One to save it into the skill as the example to match. It gives Genie One a clear target and gives you something to compare against whenever the skill changes.
Correct in conversation: say what was wrong and what you expected. Genie One edits the skill and can rerun the work straight away.
Start a fresh chat to retest: this confirms the updated skill works on its own, not just in the conversation where you fixed it.
Example: putting the skill to the test
Let's say a monthly trading brief skill has been built. Genie One pulls last month's numbers, flags anything unusual (such as a product line growing much faster than the rest), checks the team's inbox for context and writes up a short brief. On its first run, two things needed fixing. It reported on September while the month's data was still incomplete, and it flagged product lines so small that their movements didn't matter.
Both were fixed in one message. Genie One was told to only report on a month once its data is complete, and to ignore anything under $50,000 in sales. It updated the skill and reran the brief for August, showing every change in the document.
After, Genie One was then retested in a fresh chat. "Put together our trading report for last month" loaded the skill without naming it. This time Genie One reported on August, the latest complete month, and only flagged the two items big enough to matter: a jump in online marketplace returns and a spike in business sales. It searched the team's Trading Notes label in Gmail and found an email explaining each one: a retailer's free returns event and a one-off hotel order.
Govern skills with Unity Gateway
A skill should be managed like any other trusted asset. Unity Gateway Skills (Beta) give your organisation one central place to publish, discover and govern skills. When you publish a skill from Genie One, it becomes a Unity Catalog object, protected by the same permissions, tags and audit as your other data assets. You decide who can use it and who can edit it, there are no copies to keep in sync, and teams can find what already exists rather than rebuilding it.
Skills start out as your own. When one would be valuable for others in your organisation, publish it through Unity Gateway and share it with the individuals or groups who need it. They can switch it on in Genie One straight away. Anyone you give edit rights can improve the skill for everyone, and anyone else can clone it to make their own version.
A skill that runs on a regular cycle does not need you to ask for it. Create a scheduled task in Genie One and point it at the skill, for example "Run the monthly trading brief on the first working day of each month." The results arrive in a chat thread and by email. If the skill normally pauses for your review, tell it what to do on scheduled runs, when nobody is there to answer.
You don't need to open a skill to use one, but it helps to know what is inside. Every skill Genie One writes is saved in your workspace as a folder with a SKILL.md file, and you can find all of yours under Customizations > Skills in Genie One. Each skill starts with a name and a description, which is what Genie One reads to decide when the skill is relevant. Supporting text files, such as templates or a golden reference, sit alongside it and are read only when a step needs them.
The instructions that follow are plain language. This part of the trading brief skill sets out the document Genie One produces, with judgement for "What moved and why" and a watch list that must always be present.
You can edit a skill directly, or choose Edit with Genie. It’s worth reading what Genie One wrote, even when it works. Every line is an instruction it will follow each time, and reviewing it is the quickest way to spot a rule or step you didn't mean to include, or find something new to add in.
Write down your first skill recipe
Genie One already knows your business. Skills teach it how you work. Every skill you write is one less process that lives only in someone's head. Pick one task you explain more than once, map it out, build it with Genie One and test it until it comes back the way you would have done it yourself.
Text: Open Genie One and describe a task you do every month, then ask it to help you turn it into a skill. Learn more about user skills in Genie One, Unity Catalog Semantics and Unity Gateway Skills.
Get the latest posts in your inbox
原文为英文,正文由自动抓取并保留原文。
原文发布于 Databricks Blog(Teach Genie One your workflows: The ultimate guide for skills in Genie One),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。