top of page
Bell Consulting Solutions

200 Characters Decide Whether Claude Opens Your Skill

Writer: Branden Bell
Branden Bell
Sep 2
6 min read

Pricing update, September 10, 2026: expert working sessions are $250 for 60 minutes. Build in a Day starts at $1,500; 30-Day Implementation starts at $6,000. Earlier package inclusions, credits, and hourly support terms below are historical. See current services and pricing for scope and terms.


There's a line on my Claude guide that hands you two free skills to download. It's been sitting there since August 17, and in all that time I never once said what a skill is. I linked the download and moved on, which is a strange thing to do to somebody who's never heard the word.


So here's the whole thing, checked against Anthropic's own documentation on September 2, 2026. What surprised me when I went back and read it properly is how small the deciding piece is: a skill can be six files and a Python script that refuses to write a broken output, and Claude reads none of that when it's choosing whether to use it, only one line capped at 200 characters.


What a Skill Actually Is


Anthropic's support article defines it plainly:


Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks.

A folder. Inside it, one markdown file with YAML frontmatter at the top and your instructions underneath. That's the minimum viable skill, and you can write one in a text editor in about four minutes.


The Claude Code documentation gives the cleanest test for when you need one:


Create a skill when you keep pasting the same instructions, checklist, or multi-step procedure into chat, or when a section of CLAUDE.md has grown into a procedure rather than a fact.

That's it. If you're typing the same setup paragraph into a new chat every Monday, that paragraph wants to be a skill. Skills run on Free, Pro, Max, Team and Enterprise plans, they need code execution turned on in your settings, and the format is published as an open standard called Agent Skills, so what you write isn't locked to Claude.


The Line That Does All the Work


The frontmatter holds two fields. A name, capped at 64 characters. And a description, capped at 200. Here's how Anthropic's own creation guide describes that second one:


This is critical—Claude uses this to determine when to invoke your skill (200 characters maximum).

Two hundred characters is shorter than this paragraph, and it's the entire audition, because of how the loading works:


Skills work through progressive disclosure—Claude determines which skills are relevant and loads the information it needs to complete that task, helping to prevent context window overload.

Anthropic's creation guide spells the levels out. The metadata is level one. The markdown body underneath it is level two. Any reference files you drop in the folder beside it are the next step out, and Claude only walks outward when the level before it gave it a reason to.


What Sits Outside the 200 Characters


Everything else. All of it. The Claude Code docs state the boundary directly:


In a regular session, skill descriptions are loaded into context so Claude knows what's available, but full skill content only loads when invoked.

I want to be fair about why that design is right, because it's the reason skills are cheap to keep. If every skill you owned loaded in full at the start of every conversation, thirty skills would eat your context window before you typed a word. Loading descriptions only is what lets you keep a large library without paying for it.


But it does mean the failure mode is invisible. A skill that never fires doesn't error. It doesn't warn you. It sits in the folder, correct and unread, while Claude does the task the ordinary way and does it worse. And nine times out of ten the reason isn't the instructions. It's that the description didn't say the words the person actually typed.


Take the brand kit skill I give away. Its folder has six files in it, including a build script that hard fails when body text drops under a 4.5 to 1 contrast ratio. None of that helps if the description says "brand tool." It has to name the trigger: brand kit, brand identity, style guide, colours and fonts for my business, the phrases somebody would really use. I've started treating that line as a search query rather than a title, and it's the only part of a skill I now rewrite after watching it fail to fire.


1 Setting That Hides the Description Too


There's a field called disable-model-invocation, and setting it to true means only you can start the skill by typing its name. Claude can't reach for it on its own. The docs are explicit that this also pulls the description out of context entirely, so Claude no longer knows the skill exists.


That's the right call for anything with side effects. The Claude Code docs name deploy, commit and send-slack-message as examples, and the reasoning is one I'd apply to client work without hesitating: you don't want Claude deciding to deploy because your code looks ready. There's a mirror field too, user-invocable set to false, for background knowledge Claude should have but you'd never type as a command.


If you've read my post on the five permission modes, this is the same instinct one layer up, which is that you decide in advance what's allowed to start itself rather than deciding it in the moment when something is already running.


Where It Lives Decides Who Can Use It


On the desktop side, the folder path is the whole permission system:

  • Personal: the skills folder in your home directory, available across all your projects.

  • Project: a skills folder inside the project itself, so it travels with the repository and applies there only.

  • Plugin: bundled inside a plugin, live wherever that plugin is enabled.

  • Enterprise: provisioned through managed settings for everyone in the organization.


When two skills share a name, enterprise beats personal and personal beats project. In the Claude apps it's simpler: zip the folder with the folder itself as the root of the zip, then upload it under Customize and Skills.


Recording One Instead


If writing markdown isn't your thing, there's a second route worth knowing about. On Pro, Max and Team plans, inside Cowork in Claude for Mac, you can record yourself doing the task and let Claude write the skill from what it watched. It isn't in chat, it isn't on Windows, and it isn't on Free or Enterprise. A session runs about ten minutes.


The retention detail is the one I'd want a client to hear before they hit record. The video and audio aren't kept, but a set of screenshots is, inside the Cowork task, and deleting the task removes them, so close your banking tab before you start and don't say a client's numbers out loud while the microphone is on.


The Honest Wrinkle


The two documentation pages don't fully agree. The support article calls name and description required metadata and writes the filename in lowercase. The Claude Code page says every frontmatter field is optional, that only the description is recommended, and writes the filename in capitals. Same feature, two surfaces, two answers.


Neither page is wrong for its own surface, and I'd rather have two accurate pages than one vague one, but it's worth naming, because "I followed the docs" stops being much of a defense when there are two sets of them and you didn't write down which one you were reading.


What I'd Do Today


Open your last three Claude conversations and find the paragraph you pasted into all of them. Not the request, the setup. The context you retype every time about how your business works or what the output has to look like.


Make a folder. Put that paragraph in a markdown file. Then spend more time on the 200 character line above it than on the paragraph itself, because the paragraph is what the skill does and the line is whether it ever gets the chance.


Sources



The free skill I mentioned is on its build page, no email required: the Claude skill that writes your brand kit. And if you want the wider picture of which Claude surface to use for what, that's the guide to using Claude this post came out of.


Building something like this for your own business and want a second set of eyes on it? I do Claude consulting at $250 for a 60-minute expert working session, and the intro call is free.

Comments

bottom of page