2026-10-16 What Are Claude Code Skills, and What Do I Want to Build with Them?

Today, I studied Skills in Claude Code. I already had something I wanted to build, so I explored how Skills work while thinking about how they might support that idea.

I want to build a Skill that helps me gradually develop vague ideas through conversation with AI. The working name is /idea-carving: shaping an idea a little at a time.

I want to be able to start before I have a clear explanation. I should be able to ramble, change direction, and leave some questions unanswered. As parts of the idea become clearer, I want to capture them in documents and diagrams.

Finding out how a Skill could support that experience was the starting point for today’s study.

A Skill extends Claude’s capabilities through instructions written in a SKILL.md file. Claude can load it when it considers it relevant, or a user can invoke it directly with /skill-name. A Skill can also include supporting materials and scripts. Claude Code documentation

Skills are useful when the same instructions or procedures keep coming up in conversations. Their full instructions load when the Skill is used, so they do not all need to remain loaded from the beginning of every session. Claude Code documentation

In my case, there are principles I want the AI to follow when helping me develop an idea. I want enough room to explore before reaching a conclusion. I do not want a fixed questionnaire to narrow my thinking too early. I also want to avoid having an AI suggestion quietly turn into something treated as my decision.

Putting these principles into a Skill seems like a way to reduce how often I have to explain the kind of conversation I want.

I want the AI to take an active role. It should ask about gaps in my explanation, point out contradictions, and suggest possibilities I have not considered.

At the same time, suggestions and decisions need to remain distinct. An AI interpretation should stay identifiable as an interpretation, and my own thoughts should remain identifiable as mine. I want to decide which direction to explore and when something is ready to be settled.

When describing this goal, I wrote (translated from Korean):

AI should not make decisions about my ideas for me. It should be a co-designer that helps draw out and develop thoughts I have not yet managed to express.

That is the principle I most want /idea-carving to preserve.

The way it records ideas should reflect this, too. Within one idea, some parts may still be exploratory, others may be taking shape, and a few may already be settled. I want to preserve those differences instead of pushing everything toward completion at the same pace.

For example, I might have a clear picture of how someone would use a service while still being unsure about its features. In that situation, I would like to record the usage scenario and keep exploring the features. This is an example of the experience I hope to build.

I plan to save clearer pieces as Markdown documents and use tools such as Mermaid when a diagram would help explain their relationships or flow. I should not have to wait until the entire idea is finished. But I also do not want a document generated after every exchange. Documentation should happen when it helps me keep thinking.

I also want to continue across sessions. My current plan is to put the principles for conversation and documentation in the Skill, while storing ideas and decisions in separate files. I will need to design how the AI reads those records and picks up the conversation again. Creating the Skill and building a reliable way to preserve my thinking are related tasks that each need attention.

I plan to start by invoking /idea-carving manually whenever I want to use it. Claude Code supports disable-model-invocation: true, which prevents Claude from invoking a Skill automatically. That fits my preference for controlling when this process begins. Invocation controls in the official documentation

One point in the authoring guidance stood out to me: match the degree of freedom to the task. Flexible instructions suit work with multiple valid approaches, while more specific procedures or scripts suit work that needs consistency. Anthropic’s Skill authoring best practices

For my Skill, that could mean leaving the conversation open while keeping the recording format consistent. I could define how to distinguish my thoughts from AI suggestions and how to handle decisions I have confirmed. Writing those instructions will not be enough to assume the behavior works as intended; I will need to observe it in actual conversations.

I have not built and used the first version yet. Today’s work was about understanding Skills and connecting that understanding to the experience I want to create.

The official guidance also recommends observing where Claude struggles without a Skill, creating evaluation scenarios, and then adding and refining the instructions needed to address those gaps. Anthropic’s guidance on evaluation and iteration

Next, I want to build a small first version with Claude. As I use it, I will look at whether it reaches conclusions too quickly, whether its questions help me explore, and whether taking notes interrupts the conversation.

The next step I recorded was to build the Skill together, try it, and keep refining it. I want to discover where the actual experience differs from what I imagined, then shape it a little further.

Technical descriptions were checked against official documentation on October 16, 2026. The proposed behavior and structure of /idea-carving remain plans to implement and test.