Login / Register ID | EN
This page has no official English version. It was translated automatically and may contain errors. Read the original in Indonesian →
Skills di Claude: Instruksi Kerja yang Dipakai Ulang, Bukan Prompt Tempelan
Foto: Pexels
Bisnis AI

Skills in Claude: Reusable Work Instructions, Not Template Prompts

Tim Taalenta · 13 November 2026

The Same Prompt, Pasted Again Every Week

Every Monday, you paste lengthy instructions regarding the format of the weekly report: section order, table style, one-page limit. Next week, it's pasted again. If any line is missed, the result goes off track. Anthropic refers to this pattern as a clear sign: if you keep typing the same prompt in many conversations, it's time to create a skill.

What are Skills

Skills (Agent Skills) are folders containing instructions, scripts, and supporting materials that Claude loads when needed for specific tasks. The mandatory content is just one file named SKILL.md. The top of this file contains two columns, name and description. Below that, the steps for Claude are written. The same folder can be filled with templates, examples, or scripts.

Anthropic likens it to writing an orientation guide for new employees. Procedures are written once and then used every time that task arises.

How Claude Loads Them

Claude does not read all skills at once. This tiered pattern is called progressive disclosure:

  1. Initially, Claude only holds the name and description of each skill, about 100 tokens per skill.
  2. If your request matches the description, the contents of SKILL.md are read.
  3. Supporting files are opened only when necessary. Scripts are executed, and only the output enters the context.

So you can install many skills without cluttering Claude's workspace. The consequence is that vague descriptions make skills never get called.

Different from Projects, Custom Instructions, and Regular Prompts

  • Project stores background knowledge that is always loaded in every chat within that project. Skills contain procedures that are active only when relevant. Anthropic's formulation: Project says "this is what you need to know," skill says "this is how to do it."
  • Custom instructions apply to all conversations. Skills are specific to one type of task.
  • Regular prompts are suitable for one-off requests and do not carry over to the next conversation.

How to Activate It

The main requirement is one: code execution must be active. Individual users activate it in Settings > Capabilities via the option "Code execution and file creation" (in the new layout called "Create and edit files"), then set up skills in Customize > Skills. Self-created skills are uploaded as a ZIP file and remain private until you share them.

Regarding packages, the two official Anthropic pages were not uniform as of November 2026. The help page states that Skills are available in all packages including Free, while the platform documentation states that uploading self-created skills requires a paid package. Check directly in your account.

In the Team package, this feature is enabled by default at the organizational level. In Enterprise, the owner must activate it first in Organization settings > Plugins & skills, tab Policy. The owner can also provide skills for all members, and members can share skills with specific colleagues in view-only mode.

Example: Weekly Sales Report

Imagine a folder creating-weekly-reports containing:

  • SKILL.md: section order (summary, key figures, constraints, plans for next week), one-page limit, rounding rules.
  • A report template as a layout reference.
  • One example report from a week considered ideal.

The description: "Creating weekly sales reports according to the team's format. Use when the user requests a weekly report, sales recap, or weekly performance summary."

The following Monday, you just upload the data and request a recap. A similar pattern is suitable for brand guidelines: colors, tone of language, and logo rules are written once in one skill.

Skill Description Checklist

  • State what the skill does and when it is used.
  • Write from a third-person perspective ("Creating reports," not "I can help").
  • Include keywords that users commonly type.
  • Avoid general descriptions like "assisting with document matters."
  • Maximum 1,024 characters, without XML tags.
  • Keep the contents of SKILL.md under 500 lines. The rest should be split into separate files.

Install Only from Trusted Sources

Skills can execute code and install third-party packages. Anthropic mentions two main risks: prompt injection that causes Claude to act outside of intent, and data leakage. Their official advice is to install skills only from trusted sources, namely self-created or from Anthropic. Skills from other sources, including submissions from colleagues, need to be checked first: files, dependencies, scripts, and commands to contact external networks. Anthropic's example repository on GitHub also states that skills there are for demonstration and need to be tested before being used in important work.

Sources

Read also:

Ditulis oleh Tim Taalenta