Bible Network Crypto DeFi Onchain RWA AI Agent Stablecoin Chain SAFU CryptoTax DeFAI AGI Claude Me Claude Skill Claude Design Claude Cowork
Independent Media
Not affiliated with any project
Learn Claude Skills. Do Everything Better.
claudeskill-me.com
LATEST
Reviewing Superpowers: A TDD Framework That Literally Deletes Code Written Before Tests Exist  ·  CLAUDE.md, Rules, Skills, Hooks, Subagents — Which One? Anthropic's Official Seven-Method Decision Framework  ·  Reviewing Anthropic's Frontend Design Skill: Why Is It Installed 57× More Than the Runner-Up?  ·  Building Your First Skill: Turn Something You've Explained Three Times Into One Command  ·  Writing Your First System Prompt: From "You Are an Assistant" to a Role Setup That Actually Works  ·  Claude Code Adds Owner-Wildcard Marketplace Controls, Letting One Rule Allow or Block an Entire GitHub Org
practice

Building Your First Skill: Turn Something You've Explained Three Times Into One Command

30-Second Version · For the impatient
A skill doesn't save you typing — it saves you from re-figuring out how to explain the rule clearly every single time.

Full Explanation +
01 · Why did this happen?

How do I judge whether something is worth turning into a Skill — should every repeated task get packaged?

The criterion isn't "how many times has this repeated" but "is what I say each time highly similar." If something happens often but the judgment logic shifts significantly each time (say, what matters most in a copy review varies a lot by product and audience), packaging it into a fixed Skill could actually constrain the flexibility the task genuinely needs. On the other hand, if the rules you state are nearly identical every time and you're just tired of retyping them, that's exactly where a skill saves real time.

A simple self-test: think back to the last two or three times you did this task, and estimate how much of what you told Claude was repeated. If the overlap is high, that content is already a stable rule worth writing down. If it's different every time, the task genuinely needs in-the-moment judgment and doesn't fit well as a fixed document.

02 · What is the mechanism?

How long should the description field in Skill.md be — could writing it too short cause Claude to misjudge when to use it?

The official recommended cap on the description field is 1,024 characters, but length isn't the point — what matters is whether it covers both "what this Skill does" and "when to use it." Writing only something like "organizes meeting notes" that purely describes the function can cause Claude to miss triggering it when the context isn't obvious enough. A more reliable approach lists specific trigger contexts too — something like "use when the user mentions meeting notes, action items, or asks to organize meeting highlights" — including keywords the user might actually say.

Also worth noting: write in third person ("organizes XX" rather than "I can help you organize XX"), because the description gets injected directly into the System Prompt, and inconsistent point of view can confuse Claude's judgment. If you find in actual use that a skill should have triggered but didn't, the first place to check is usually whether the description states the trigger context specifically enough.

03 · How does it affect me?

Should rules in a Skill be written as bullet points or full paragraphs? Is there a more effective format?

This depends on the nature of the rules themselves, and the concept of "Degrees of Freedom" is useful here. If rules are independent of each other and the judgment resembles a checklist ("check for an owner," "check for a deadline"), bullet points are clearest — Claude can go through them item by item. If a rule requires some reasoning logic with sequence or conditional branches ("first determine whether this is an action item, and if so, decide which category it belongs in"), numbered steps or a workflow description conveys the logical relationship better than bullets.

Another practical technique is providing concrete input-output examples rather than only describing rules in words. For instance, instead of writing "don't leave the owner field blank," demonstrate it directly with a specific case: "Input: 'discussed the marketing budget, everyone agreed it needs replanning' → Output: owner field marked 'unassigned,' since no one was designated." Examples often convey the exact standard you want more clearly than an abstract description can, especially for rules with a somewhat fuzzy boundary.

04 · What should I do?

If I followed this process to write a Skill, and after using it a few times found it's not quite accurate in some cases, how do I fix it efficiently?

A more efficient fix isn't guessing where the problem might be from memory — it's taking the actual case where a gap occurred and pinpointing exactly which step judged wrong. For example, if you notice Claude misclassified a piece of pure information-sharing as an action item, first check whether the existing Skill.md rule "pure information-sharing isn't listed" was written clearly enough, or whether this particular content happened to fall into a gray area the rule didn't clearly cover. If it's the latter, you need to add a more specific judgment criterion or example — not rewrite the whole rule.

This refinement process can lean on Claude itself: paste the current SKILL.md content alongside the specific case where the judgment went wrong, and directly ask it to help analyze which part of SKILL.md might not be clear enough to cause this misjudgment. This usually gets you a concrete, actionable suggestion faster than redesigning the rules from scratch on your own.

Full Content +

There's a simple check for whether something is worth turning into a Skill: think back over the past week and ask whether there's anything you've already explained to Claude at least two or three times, with roughly the same content each time. If so, that's a good candidate for packaging into a Skill — instead of re-describing it every time, write that explanation into a document once, and Claude applies it automatically whenever a matching task comes up.

Here's a walkthrough using a common scenario: say you organize scattered meeting notes into a standardized action-item list for your team every week, and every time you have to re-explain to Claude what fields to use and how to categorize things.

Step one: do the task normally first, without writing a skill

Before writing a skill, complete the task once with Claude the normal way, and jot down every rule you actually stated along the way. For organizing meeting notes, you might have said things in the conversation like: "each action item needs an owner, a deadline, and a current status field"; "an item with no clear owner should be flagged 'unassigned'"; "pure information-sharing with no follow-up action shouldn't go into the action-item list." These rules you stated on the fly in conversation are exactly what the skill should contain.

Step two: write these rules into SKILL.md

The most basic form of a skill is a Markdown file containing YAML frontmatter (recording its name and description) plus a body explaining what it does. Continuing the example above, it looks roughly like this:

Example SKILL.md:

"---
name: organizing-meeting-action-items
description: Organizes meeting notes into a standardized action-item list. Use when the user mentions meeting notes, action items, or asks to organize meeting highlights.
---

## Organizing rules

Each action item includes three fields: owner, deadline, current status.

An item with no clear owner should be marked "unassigned" in the owner field, not left blank or marked "unspecified."

Pure information-sharing content with no action required from anyone should not go into the action-item list — organize it separately into a "meeting notes" section instead.

Output as a table, with a fixed column order: action item, owner, deadline, status."

The most important part of this skill description is the description field — it determines when Claude proactively thinks to use this skill, so it needs to specifically state what the skill does and when it applies, rather than something vague like "organizes meeting notes."

Step three: test it for real, and refine based on results

Once written, hand Claude an actual set of meeting notes as a test and see whether it really follows these rules. If you notice a rule got missed (the owner field is still blank instead of showing "unassigned"), that means that rule wasn't stated clearly or prominently enough in SKILL.md — go back, adjust the wording or move the important rule further up, and test again. This test-find-a-gap-fix cycle usually takes two or three rounds before it stabilizes.

What This Means for Your Work

What a skill actually saves isn't typing time — it's the mental effort of having to re-figure-out how to describe a rule every single time. Once written as a skill, that piece of judgment logic becomes a reusable asset, one you can also share with the rest of your team. The barrier to starting is low — you don't need to write a perfect version on the first try. Just work through the task once in a normal conversation, write down the rules as you state them, save it as a Markdown file — that's already a minimally usable skill, and you can fill in the details later as needed.

Diagram
打造第一個 Skill 的三步驟流程圖呈現「正常做一次任務並記錄規則→寫成 SKILL.md→實際測試並修正」的三步驟循環,並以虛線箭頭標示步驟三發現問題後回到步驟二修正、需重複兩三輪才會穩定的疊代過程。Building Your First SkillStep 1Do the task oncewrite down each ruleStep 2Write SKILL.mdname + description + rulesStep 3Test with real datafind gaps, refineRepeat 2-3 rounds until stableClaude Skill Me · claudeskill-me.com
Feel free to share. Please credit the source.
Ask a Question
Please enter at least 10 characters
Related Articles
Reviewing Superpowers: A TDD Framework That Literally Deletes Code Written Before Tests Exist
reviews · Aug 15
CLAUDE.md, Rules, Skills, Hooks, Subagents — Which One? Anthropic's Official Seven-Method Decision Framework
advanced · Aug 15
Reviewing Anthropic's Frontend Design Skill: Why Is It Installed 57× More Than the Runner-Up?
reviews · Aug 15
Writing Your First System Prompt: From "You Are an Assistant" to a Role Setup That Actually Works
prompt-examples · Aug 14
Related News
More Related Topics