探索预构建技能

摘要

本课程介绍Anthropic提供的预构建技能生态系统，包括文档处理技能（Excel、PowerPoint、Word、PDF）和技能创建工具。讲师演示如何通过skill-creator以编程方式创建新技能，如何将现有技能从CSV迁移到BigQuery数据源，以及如何结合品牌指南技能与内置PowerPoint技能构建自动化演示生成工作流。

要点

• Anthropic预构建技能分为两类：文档技能（Excel、PowerPoint、Word、PDF）始终内置启用；示例技能（如skill-creator）需手动开启
• skill-creator是一个元技能，能通过编程方式创建新技能，自动生成文件结构、YAML frontmatter和验证脚本
• 技能可与MCP服务器集成，实现从外部数据源（如BigQuery）直接获取数据，无需手动上传文件
• 演示完整工作流：使用BigQuery MCP服务器获取营销数据，结合自定义品牌指南技能和内置PowerPoint技能自动生成演示文稿
• 技能创建遵循最佳实践：明确的步骤说明、具体的示例、避免模糊的日期范围、保持必要的上下文和约束

视频信息：Exploring Pre-Built Skills

中文翻译

在第一节课中,Claude AI 使用 Excel 技能创建了展示营销结果的电子表格。Excel 技能是 Anthropic 的预构建技能之一,其中还包括 PowerPoint、Word 和 PDF 技能,以及一个技能创建技能(Skill Creator)。让我们来看看这些技能。

现在我们已经了解了技能如何融入整个 AI 生态系统,让我们来看看你可以直接在 Claude AI 和 Claude Desktop 中开箱即用的一些预构建技能,以及你可以使用 Claude Code 等工具自行安装的技能。在这个位于 github.com/anthropic/skills 的存储库中。让我们看看 skills 文件夹,看看我们有哪些内置技能。所有这些技能都已经准备好用于生产环境。实际上,我们在之前的课程中已经看到了这个 Excel 技能的用例。值得注意的是,这份由 Anthropic 创建的技能列表实际上被分成了两个不同的部分。针对 Microsoft Docs、PDF、PowerPoint 和 Excel 的技能被称为文档技能。这些技能是内置的,并且在 Claude AI 等工具中始终被使用。其余的这些技能是我们创建的示例,你可以在 Claude 中开启或关闭它们,但默认情况下,除了 skill-creator(技能创建器)之外,它们都是关闭的。

让我们首先分析一下 PowerPoint 技能。我们可以看到,就像其他结构一样,我们有一个 SKILL.md 文件以及可供引用的其他文件和文件夹。在这个 SKILL.md 中,我们要有包含名称和描述的相同 YAML Frontmatter(前置元数据)。你在这里看到的是 GitHub 如何渲染这个 Markdown 文件,但底层代码看起来与我们之前制作的非常相似。如果你熟悉 Markdown 文件,你可以这样查看它。我会切换回预览模式,因为它看起来稍微好一点。当我们查看这个技能如何工作以及它做什么时,我们看到了一个概览。用户可能会要求创建、编辑、分析 PowerPoint 文件的内容。这就是它的样子,这就是你阅读它的方式。如果需要完成特定任务,会有底层脚本去执行。请记住,这些脚本不是开箱即用的。它们仅在必要时才被加载和执行。

我们可以用 PowerPoint 演示文稿做很多事情,比如颜色、排版,正如你可以想象的那样,这就是我们如何开始制作看起来更好、更像真实世界演示文稿的东西。我们有设计原则、必要的要求,以及当用户没有指定颜色时,我们可以让 Claude 从中挑选的调色板选择。这个 SKILL.md 文件很长,因为我们可以用 PowerPoint 演示文稿做很多事情。但我们在本节课后面将看到的是如何实际使用这个技能,将现有数据转化为一个美观的演示文稿。

我想向你展示的下一个技能是一个有点元概念(Meta Idea)的东西,那就是 skill-creator(技能创建器)。技能创建器是一个旨在为你以编程方式创建技能的技能。与其必须从头开始做事情并创建必要的文件和文件夹结构,不如让技能创建器为你做这件事。让我们来看看 SKILL.md 文件,看看这里发生了什么。与我们的其他技能类似,我们有一个名称和一个描述。实际上我还要看一下这里的底层代码,因为它更容易理解。我们在 SKILL.md 文件中指定了什么是技能,它提供了什么,然后我们包含了一些与技能相关的最佳实践。我们将在下一节课中深入探讨这些最佳实践。但你可以想象,当 Claude 为你以编程方式创建技能时,我们要利用其中一些最佳实践。

当我们查看技能创建过程时,我们在这里列出的步骤非常明确。因为我们想使用这个技能来创建一个可预测的工作流,所以我们要非常明确地说明步骤是什么,如何遵循它们,以及只有在存在某种原因时才跳过什么。我们从具体的例子开始,我们规划可重用的技能内容,在这里你可以开始看到非常有用的例子,以便 Claude 在你想创建技能时进行模式匹配。当我们开始初始化技能时,这里我们要运行底层 Python 脚本来执行必要的任务。让我们看看这些脚本是做什么的。在 scripts 文件夹中,我有三个 Python 文件。一个用于初始化技能并提供底层文本的脚本,一个用于打包该技能的 Python 文件,还有一个用于验证该技能的脚本。

让我们看看这个底层代码是如何初始化技能的。我们采用一个现有的模板,其中包含一些 YAML Frontmatter 以及一些占位符和待办事项(To-Dos),然后根据传入的数据将其填充。这个底层脚本允许我们在制作技能时创建必要的文本文件。一旦我们生成了必要的文件,我们就可以将其打包。在这里你可以看到我们引入了必要的模块来压缩我们所需的技能,并确保我们在正确的文件夹和文件结构中执行此操作。最后,我们还有一个脚本来执行技能验证。确保 SKILL.md 存在,验证一些 YAML Frontmatter,并确保我们放入文件夹和文件中的内容是正确的。我们将利用这个 skill-creator 技能来获取我们现有的内容,并将其打包成一个可重用的模块化脚本。

现在,让我们把话题转回 Claude,看看如何将内置技能、我们自己的技能以及带有 MCP 服务器的可预测工作流结合在一起。回到 Claude,让我们去看看并确保我们启用了正确的技能以及它们的位置。回到设置中,在"功能"(Capabilities)里面。我们之前看到过,我们可以在这一部分创建技能。我想向你展示的是我们拥有的示例技能,这看起来应该很熟悉。这就是我们在 GitHub 上看到的。默认情况下,这些技能是关闭的。如果我们想开启它们,这完全没问题。默认开启的技能是我们刚刚看到的 skill-creator。值得注意的是,虽然 skill-creator 在创建底层技能和必要结构方面非常有效,但我们仍然必须对我们提供的提示词以及我们要制作的技能所输入的数据保持目的性。

我们现在要做的是将所有这些关于技能、MCP 和提示词的想法整合在一起。首先,我们要修改之前创建的用于分析活动的技能。不再使用 CSV 作为数据,而是使用 BigQuery。如果你不熟悉,BigQuery 是 Google 提供的数据存储服务,为了引入必要的工具和上下文来与 BigQuery 协作,我们将连接一个 MCP 服务器。所以我们将使用 skill-creator 技能来修改我们之前的营销分析技能以使用 BigQuery。然后,我们将使用 skill-creator 创建另一个技能。这将用于品牌指南。我们将包含一个指定指南和徽标的文件,我们将为自己构建另一个技能来执行该任务。最后,我们将采用我们用于提取和分析数据以及利用品牌指南的两个技能,并将它们与用于创建 PowerPoint 演示文稿的内置技能相结合,以创建一个利用提示词、技能和模型上下文协议的工作流。

在开始之前,你可能想知道我们之前看到的 Excel 和 PowerPoint 以及其他文档技能在哪里。这些是内置于 Claude AI 中的,不是可以开启和关闭的东西。考虑到这一点,让我们开始这个工作流。在我们将分析营销活动的技能修改为使用 BigQuery 之前,还要注意我们在这里使用 Claude Desktop 连接到本地 MCP 服务器以利用 BigQuery。所以让我们看看那个 BigQuery 服务器是如何配置的。我将前往"设置"、“开发者”。在这里,我们可以查看底层命令和参数、特定项目的环境变量以及我的凭据所在的位置。在这个例子中,我们不必使用 BigQuery,你可以使用数据库、一些外部数据存储,但我们只是想展示技能和 MCP 服务器协同工作的样子。如果你有兴趣查看那个底层配置文件,它就是这个样子的。在这个配置文件中,我们指定了我们想要连接的服务器以及 Claude Desktop 启动时运行的底层命令。

考虑到这一点,让我们继续修改我们之前的技能,现在使用 BigQuery 而不是 CSV 进行数据访问。为了确保这能正常工作,让我们首先让 Claude 列出 BigQuery 中存在的表。这将利用我们拥有的 MCP 服务器。我们将允许这样做,我们应该会得到表列表。在这种情况下,我们只有一个。所以在这里我们可以看到有一个名为 marketing 的数据集,其中包含一个表。现在我们要让 Claude 向我展示表的架构(Schema)。希望 Claude 能识别出那个小的拼写错误,我们就可以开始工作了。在这里我们指定了表的样子。这看起来很棒。当我们继续更新我们的 analyzing-marketing-campaign 技能时,我们将利用这个架构。

我们现在要做的是让 Claude 更新我们的 analyzing-marketing-campaign 技能,以便不再上传 CSV,而是从 BigQuery 拉取数据。我们指定来自 BigQuery 表的数据,特别是我们刚刚在上面看到的架构。由于我们都在同一个对话中,Claude 应该可以毫无问题地查看架构是什么。我们为此指定了一些要求,就像在我们现有的技能中一样,我们要确保对预算重新分配规则的引用不会被修改。就像我们之前谈到的那样,技能创建器技能非常有帮助且高效,但我们仍然需要给出必要的上下文。请注意这里,它要做的第一件事是分析必要的技能结构,并使用我们的技能创建器技能来修改现有的技能并遵循最佳实践。

我们要现在继续创建一个带有新 SKILL.md 文件的更新技能。在这里我们可以开始看到一些感觉类似于我们之前技能的东西。但是改为添加 BigQuery 而不是 CSV 上传。在底层,我们使用文件系统和 bash 工具为我们创建必要的文件和文件夹结构。我们在这里可以看到的是,我们不再使用 CSV,而是使用 BigQuery,并且我们遵循了将 MCP 服务器与技能一起使用的最佳实践,即我们指定服务器和工具的名称。skill-creator 正在遵循最佳实践来获取我们要修改的现有技能。正如我们指示 skill-creator 那样,当我们指定我们所需的输入时。我们现在正在实践中看到这一点。最好的做法是不要使用模糊的日期范围或整个范围,所以我们要求用户澄清,当我们展示查询示例时,我们要指定日期范围。所以我们放入的一些工具和要求在我们更新此技能时被直接应用了。所以我们的技能看起来状态良好。为了确保这被保存到后续对话中,让我们继续复制这个技能。

现在我们要换个档位,为品牌指南创建一个新技能,我们将与这个技能一起使用它来创建一个引人注目的数据驱动型 PowerPoint 演示文稿。所以让我们开始一个新的聊天,我们要让 Claude 根据我们上传的文件创建一个品牌指南技能。我要做的第一件事是上传一个包含我的品牌指南以及演示文稿中要使用的一些徽标的文件。在我们继续创建这个技能之前,让我向你展示一下这些品牌指南的样子。我有一个调色板、辅助颜色、排版。Claude 知道如何设计东西,但技能真正大放异彩的地方在于你可以确切地告诉 Claude 你希望如何为你的公司做事。徽标、颜色、字体,很棒的例子。

现在让我们根据这些文件创建一个技能,我们可以将其应用于未来的演示文稿和文档。我们将再次看到 skill-creator 技能的实际应用。我们利用现有的工具和技能来使用最佳实践以及指南和徽标,以制作一个可重复且可移植的技能。我们将分析其他现有技能,看看它们使用什么模式,并确保我们正在创建的这个新技能可以与它们互补。这非常有价值,因为我们将把它与 PowerPoint 演示文稿一起使用。现在我们已经很清楚需要做什么,让我们运行我们之前看到的 init_skill Python 脚本。这将创建底层技能,现在我们可以开始将我们的资源添加到技能的 assets 文件夹中。

我们将开始看到颜色填充、强调色、字体、排版。稍后,我们将拥有一个技能,当我们需要进行设计时,我们可以开始将其添加到所有未来的对话中。我们的徽标正在被拉入,Word 文档和 PDF 已指定,演示文稿布局正是我们要的样子。技能创建器已运行完毕。在这里,我们创建了一个 SKILL.md 文件,遵循包含名称和描述的最佳实践,以及包含我们需要的数据和徽标的底层文件夹。我们还需要做一个步骤,以确保这被添加到未来的对话中。为了确保这被保存到后续对话中,让我们继续复制这个技能。一旦完成,我们应该会在我们创建的技能列表中看到这个技能。

现在我们已经更新了我们的技能以从 CSV 转移到 BigQuery,并为我们的品牌指南创建了一个新技能,让我们将其结合起来构建一个工作流,连同内置的 PowerPoint 演示文稿技能,首先分析我们的数据,然后生成演示文稿。所以我们将首先在 BigQuery 中分析不同周的营销数据,看看每个渠道的表现如何。然后基于该数据,使用我们的品牌指南生成演示文稿。让我们看看这看起来像什么。首先,我们将继续阅读相关的技能文件。这包括我们的营销活动分析,也将包括我们的 BigQuery 指南。我们将继续确保我们有正确的 PowerPoint 演示文稿技能以及用于样式的品牌技能。

在底层的 PowerPoint 演示文稿技能中,有关于演示文稿创建的额外文档。首先,我们将继续从 BigQuery 开始。我们将查询必要的内容。我们可以查看并看到正在编写的底层 SQL,就像我们之前看到的那样,那是我们正在寻找的日期范围。现在我们有了数据,我们将使用这些指标来继续生成 PowerPoint 演示文稿。我们将按照我们在品牌风格中建议的样式这样做,并将其转化为 PowerPoint 演示文稿。可以看到这里正在为我们的幻灯片编写底层 CSS 和 HTML。然后我们将依靠内置技能来创建底层演示文稿。

现在我们已经有了正确的 HTML 文件,让我们继续创建我们的演示文稿。这里我们使用原生 PowerPoint 技能并编写必要的代码来创建演示文稿。我们可以看到,即使存在特定问题,模型也会回过头来编辑任何必要的内容,并依靠确切的工作流,不仅运行必要的代码,还要验证需要做什么。模型具有回溯和遵循特定模式的能力,这使我们能够创建不带有我们需要立即纠正的内置问题的演示文稿。所以我们看到 Claude 已经完成了它的验证,幻灯片看起来很棒。现在它将继续生成那个底层 PowerPoint 演示文稿,我可以在 Google Drive 中打开它并作为 Google Slides 使用,或者我可以直接下载。

我们可以看到,我得到了一些看起来非常不错的幻灯片,带有颜色、字体、徽标以及我想为我的特定公司获得的所有内容。我们有效率分析、漏斗分析以及强调需要审查的内容和做得很好的内容的执行摘要。我可以下载这个演示文稿,我可以继续基于它进行构建,并再次在 Google Drive 中打开它与队友分享。我可以继续使用提示词并处理这个演示文稿。但我们在这里看到的是一个由内置技能创建的底层 PowerPoint 演示文稿,结合了我们制作的两个技能,以及从 BigQuery 拉取数据的 MCP 服务器。在下一节课中,我们将探索一些关于创建技能的最佳实践,并看看我们创建的另外两个自定义技能,看看我们是否遵循了最佳实践。

English Script

In the first lesson, Claude AI used the Excel skill to create spreadsheets displaying the marketing results. The Excel skill is one of Anthropic’s pre-built skills, which also include a PowerPoint, Word, and PDF skill, as well as a skill creation skill. Let’s take a look at those.

Now that we’ve seen how skills fit in the entire AI ecosystem, let’s take a look at some of the pre-built skills that you can use out of the box with Claude AI and Claude Desktop, and that you can install yourself with tools like Claude Code. Inside of this repository that lives at github.com/anthropic/skills. Let’s take a look at the skills folder and see which built-in ones that we have. All of these are ready for production usage. And we actually saw in a previous lesson the use case of this Excel skill. It’s important to note that this list of skills, while created at Anthropic, is actually bucketed into two different sections. The skills for Microsoft Docs, PDFs, Power Points, and Excel are known as document skills. These are built in and always used in tools like Claude AI. The remainder of these skills are examples that we’ve created that you can toggle on and off in Claude, but by default with the exception of skill-creator are toggled off.

Let’s first start by analyzing the PowerPoint skill. We can see just like other structures that we have a SKILL.md file as well as other files and folders to reference. Inside of this SKILL.md, we have that same YAML Frontmatter that includes the name and the description. What you’re seeing here is how GitHub is rendering this markdown file, but the underlying code looks very similar to what we’ve made before. You can view it this way if you’re familiar with markdown files. I’ll switch back to the preview because it looks a little bit nicer. When we take a look at how this skill works and what it does, We’ve got an overview. The users may ask to create, edit, analyze contents of a PowerPoint file. Here’s what it looks like, here’s how you read it. And if there are particular tasks that need to be done, there are underlying scripts to go ahead and execute. Remember, these are not executed right out of the box. These are only loaded and executed when necessary.

There’s quite a bit that we can do with PowerPoint presentations, colors, typography, as you can imagine, this is how we can start to make things that look nicer and look more like real-world presentations out of the box. There are design principles that we have, requirements that are necessary, and color palette selections that we can have Claude pick from when the user does not specify them. This SKILL.md is quite long, as there is quite a bit that we can do with PowerPoint presentations. But what we’re going to see later in this lesson is how to actually use this skill to take existing data and turn it into a beautiful looking presentation.

The next skill I want to show you is a little bit of a meta idea here, and that’s called the skill-creator. And the skill creator is a skill that serves the purpose of programmatically creating skills for you. Instead of having to do things from scratch and create the necessary files and folder structure, the skill creator can do that for you. Let’s take a look at the SKILL.md file and see what’s happening here. Similar to our other skills, we have a name and a description. And I’m actually going to take a look at the underlying code here, since it’s a little bit easier to follow. We specify in this SKILL.md file what skill is, what it provides, and then we include some of the best practices associated with skills. We’re going to dive into those best practices in the next lesson. But you can imagine when Claude is programmatically creating skills for you, we want to leverage some of these best practices.

When we take a look at the skill creation process, we’re extremely explicit with the steps that we have here. Since we want to use this skill to create a predictable workflow, we want to be extremely explicit with what the steps are, how to follow them, and what to skip only if some reason exists. We start with concrete examples, we plan reusable skill contents, and here you can start to see examples that are very helpful for Claude to pattern match when there’s a skill you’d like to create. When we start initializing the skill, here we’re running underlying Python scripts to perform the task necessary. Let’s take a look at what those scripts do. Inside of the scripts folder, I have three Python files here. A script to initialize the skill and provide the underlying text, a Python file to package that skill, and then a script to validate that skill.

Let’s take a look at what this underlying code does to initialize a skill. We take an existing template that we have with some YAML Frontmatter and some placeholders and to-dos, and we fill that in based on the data that is coming in. This underlying script allows us to create the necessary text files when making our skills. Once we’ve generated the necessary files, we can package that up. Here you can see we’re bringing in the necessary modules to zip our skill necessary and make sure that we’re doing this in the right folder and file structure. Finally, we have one last script to perform a validation of our skill. make sure that a SKILL.md exists, validate some of the YAML Frontmatter, and make sure that what we put inside of our folder and files is correct. We’re going to be leveraging this skill-creator skill to take existing content that we have and package it up into a reusable and modular script.

Now let’s go ahead and shift gears back to Claude and see how to put together built-in skills, our own skills, and a predictable workflow with an MCP server. Back in Claude, let’s go and take a look and make sure that we have the correct skills enabled and where those live. Back in settings, inside of capabilities. We saw previously, we can create skills in this section. What I want to show you are the example skills that we have and this should look pretty familiar. This is what we saw on GitHub. By default, these skills are turned off. If we want to toggle them on, we can absolutely do so. The skill that is toggled on by default is the skill-creator that we just saw. It’s important to note that while the skill-creator is extremely effective at creating underlying skills and structure necessary, we still have to be intentional about the prompt that we provide and the data that goes in to the skill that we’re going to make.

What we’re going to do now is put all of these ideas around skills, MCP, and prompting together. First, we’re going to modify our previous skill that we created for analyzing campaigns. to not use a CSV for data, but instead BigQuery. If you’re not familiar, BigQuery is a data store powered by Google, and in order to bring in the necessary tooling and context to work with BigQuery, we’re going to connect an MCP server. So we’re going to use the skill-creator skill to modify our previous marketing analyzing skill to use BigQuery. We’re then going to use skill-creator to create another skill. This will be for the purpose of brand guidelines. We’ll include a file that specifies the guidelines as well as logos, and we’ll build for ourselves another skill to perform that task. Finally, we’ll take our two skills that we used to extract and analyze data and to leverage brand guidelines and combine them with a built-in skill for creating PowerPoint presentations to create a workflow that makes use of prompting, skills, and the model context protocol.

Before we jump in, you might be wondering where the Excel and PowerPoint and other document skills that we saw before live. These are built in to Claude AI. and are not things that can be toggled on and off. So with that in mind, let’s start this workflow. Before we modify our analyzing marketing campaign skill to use BigQuery, Let’s also make a note that we’re using Claude desktop here to connect to a local MCP server to leverage BigQuery. So let’s take a look at how that BigQuery server is configured. I’m going to head over to Settings, Developer And here, we can take a look at the underlying command and arguments and environment variables for the particular project and where my credentials live. For this example, we don’t have to use BigQuery, you can use a database, some external data store, but we just want to showcase what it looks like with skills and MCP servers working together. And if you’re interested in seeing that underlying config file, here’s what it looks like. In this config file, we specify the servers we want to connect to and the underlying commands to run when Claude Desktop starts.

With that in mind, let’s go ahead and modify our previous skill to now use BigQuery instead of CSVs for data access. To make sure this is working correctly, let’s first ask Claude to list the tables in BigQuery that exist. This is going to make use of the MCP server that we have. We’re going to allow this and we should get back the list of tables. In this case, we only have one. So here we can see there’s a data set called marketing that contains a single table. Now we’re going to ask Claude to show me the schema of the table. Hopefully Claude can pick up that small spelling mistake and we should be in business. Here we’re specifying what the table looks like. And this looks great. And we’re going to make use of this schema when we go ahead and update our analyzing-marketing-campaign skill.

What we’re going to do now is ask Claude to update our analyzing-marketing-campaign skill so that instead of a CSV upload, we pull from BigQuery. We specify the data from the BigQuery table, specifically the schema that we just saw above. Since we’re all in one single conversation, Claude should have no problem taking a look at what the schema is. We’re specifying some requirements for this, and just like in our existing skill, we want to make sure that the reference to our budget reallocation rules does not get modified. Like we spoke about earlier, the skill creator skill is extremely helpful and efficient, but we still need to give the context necessary. Notice here, the first thing it’s going to do is analyze the necessary skill structure and use our skill creator skill to modify the existing skill and follow best practices.

We’re going to go ahead now and create the updated skill with a new SKILL.md file. Here we can start to see something that feels similar to our previous skill. But instead adding BigQuery instead of CSV uploads. Under the hood, we’re using the file system and bash tools to create the necessary file and folder structure for us. What we can see here is instead of using a CSV, we’re using BigQuery and we’re following the best practice of using MCP servers with skills where we specify the server and the name of the tool. The skill-creator is following best practices to take our existing skill and modify it. So as we instructed skill-creator, when we specified our required input. We’re seeing this in practice right now. It’s best practice not to use an ambiguous date range or the entire range, so we ask the user to clarify, and when we show an example of querying, we’re specifying a date range. So some of the tools and requirements that we put in are being directly applied when we update this skill. So our skill looks like it’s in great shape. In order to make sure this is saved to subsequent conversations, let’s go ahead and copy this skill.

Now we’re going to shift gears and create a new skill for brand guidelines that we’ll use alongside this skill to create a compelling data-driven PowerPoint presentation. So let’s go ahead and start with a new chat and we’re going to ask Claude to create a brand guideline skill from files that we upload. The first thing I’m going to do is upload a file with my brand guidelines as well as some logos to be used in the presentation. Before we go ahead and create this skill, let me just show you what these brand guidelines look like. I’ve got a color palette, supporting colors, typography. Claude knows how to design things, but where skills really shine are where you can tell Claude exactly how you want things done for your company. logos, colors, fonts, great example.

Now let’s go ahead and create a skill from these files that we can apply to future presentations and documents. What we’re going to see here is the skill creator skill in action again. We’re leveraging the existing tooling and skills that we have to use best practices as well as the guideline and logos to make a skill that is repeatable and portable. We’re going to analyze other existing skills to see what patterns they use and make sure that this new skill we’re creating can complement them. And this is extremely valuable since we’re going to be using this with PowerPoint presentations. Now that we have a good idea of what needs to be done, let’s run that init_skill Python script that we saw before. This will create the underlying skill, and now we can start adding our assets to the skill’s assets folder.

We’re going to start to see colors populate, accent colors, fonts, typography. And in a bit, we’ll have a skill that we can start adding to all future conversations when there’s design that we need done. Our logos are being pulled in, Word documents and PDFs are specified, and presentation layouts are the way that we want them to be. The skill creator has finished running. And here we have a SKILL.md file that’s been created, following best practices with a name and a description, as well as underlying folders with the necessary data and logos that we need. There’s one more step we need to do to make sure that this gets added to future conversations. In order to make sure this is saved to subsequent conversations, let’s go ahead and copy this skill. Once this is done, we should see this skill in the list of skills that we’ve created.

Now that we’ve updated our skill to move from CSVs to BigQuery, and created a new skill for our brand guidelines, let’s combine that to build a workflow alongside the built-in PowerPoint presentation skill to first analyze our data and then generate a presentation. So we’re going to first analyze our marketing data for a different week in BigQuery to see how each channel is doing. And then based on that data, generate a presentation with our brand guidelines. Let’s see what this looks like. First, we’re going to go ahead and read the relevant skill files. This includes our marketing campaign analysis and will include our BigQuery guidelines as well. We’re going to go ahead and make sure we have the correct PowerPoint presentation skill as well as our brand skill for styling.

Inside of the underlying PowerPoint presentation skill, there’s additional documentation for presentation creation. First, we’re going to go ahead and start with BigQuery. We’re going to query what’s necessary. We can take a look and see the underlying SQL that’s being written and like we saw before, that date range that we’re looking for. Now that we have the data, we’re going to use these metrics to go ahead and generate a PowerPoint presentation. We’re going to do so with the styling that we’ve advised in our brand style and turn this into a PowerPoint presentation. can see here the underlying CSS and HTML being written for our slides. And then we’re going to lean into the built-in skill for creating the underlying presentation.

Now that we’ve got the right HTML files, let’s go ahead and create our presentation. Here we’re using the native PowerPoint skill and writing the necessary code to create the presentation. We can see here even when there are particular issues, the model will go back edit anything necessary and lean on the exact workflow, not only for running code necessary, but validating what needs to be done. This ability that the model has to backtrack and follow particular patterns allows for us to create presentations that don’t come with built-in issues that we need to immediately then correct. So we’re seeing that Claude’s done its verification, the slides look great. Now it’s going to go ahead and generate that underlying PowerPoint presentation, which I can open up in Google Drive and use as Google Slides, or I can download directly.

We can see here, I’ve got some really nice looking slides with the colors, fonts, logos, and everything that I want for my particular company. We have our efficiency analysis, funnel analysis, and the executive summary that highlights what needs review and what’s doing quite well. I can download this presentation, I can continue to build off of it, and again, open it up in Google Drive to share with teammates. I can continue prompting and working with this presentation. But what we’re seeing here is an underlying PowerPoint presentation created from a built-in skill, combined with two skills that we’ve made alongside an MCP server pulling in data from BigQuery. In the next lesson, we’ll explore some of the best practices around creating skills and take a look at two other custom skills that we create and see if we’re following the best practices.