You add one folder for a single bug, then open Claude Code with it waiting in the background. A skill is a folder of instructions the model opens when it decides it needs them. The catalogue measured 247,505 of them. Most preload a short description before you ask for help.
The name arrives before the instructions
The context is everything the model can see while it answers you. A token is the unit your AI plan bills by, about three quarters of a word.
Before the model decides a skill is relevant, only its name and description enter the context. That is what the every-session figure counts. 239,735 skills load that entry every session. Another 7,770 wait until Claude Code calls for them.
The description does two jobs. It helps Claude Code notice the skill and sets the initial cost. The full instructions arrive only when the skill opens.
The preload price of a Claude Code skill
Among skills loaded every session, the median preload is 58 tokens. The 90th-percentile preload is 141 tokens, so nine out of ten stay at or below that level.
The largest listed preload is 2,606 tokens in role-play. At 20 sessions a day, the 58-token median becomes 1,160 tokens across a day.
A 58-token preload is smaller than a short email, but it repeats whether the task needs the skill or not. On-demand skills avoid that repeated introduction and pay their larger cost when Claude Code opens them.
Security signals across the skill shelf
Every mod in the catalogue goes through the same scan. 234,827 skills received grade A, 6,959 received B, 5,708 received C or D, and 11 received E or F.
The scan flags 18,409 skills for making network calls. It flags 2,952 for asking for root access and 2,501 for recursive force deletion.
Three C-grade examples are here-now, launch, and advisor-orchestrator-worker. A finding points you to the line that triggered it. It is a prompt to inspect the file, not a judgement on the whole skill.
Six skills for common coding jobs
Every session, the name and description arrive before a request.
The original is the first copy of a file we saw; everything matching it afterwards is a copy.
- systematic-debugging: A step-by-step method for investigating bugs, failed tests, and unexpected software behaviour before changing code. 21 tokens every session, 284 lines, original, 109 copies.
- local-ai-agents: A guide for building AI agents on your own computer with local models, tools, document search, and local servers. 200 tokens every session, 97 lines, original, 0 copies.
- next-cache-components-adoption: A workflow for enabling Next.js Cache Components and resolving the route errors that follow. 95 tokens every session, 250 lines, original, 4 copies.
- playwright-cli: drives a browser from the command line to open pages, fill forms and click through them. 19 tokens every session, 426 lines, original, 29 copies.
- imagegen: A guide to generating or editing pixel-based images for websites, games, mockups, and logos. 113 tokens every session, 316 lines, original, 22 copies.
- cordis-plugin-development: Instructions for creating, modifying, debugging, or diagnosing Cordis plugins, tools, events, and interface pieces. 80 tokens every session, 421 lines, original, 9 copies.
When a hook fits better
A hook is a shell command your agent runs for you when something happens. Use a hook when you need an automatic event response, such as running a test before each commit.
A skill supplies instructions for a task Claude Code chooses during its work. It cannot replace an event-triggered command. Use the hook for automation and the skill for guidance.
Questions people ask
Do I pay for a skill before Claude Code needs it? Across these skills, the median body once one is open is 1,404 tokens. That larger charge arrives when Claude Code opens the skill, not merely because you installed it.
How many skills should I keep installed? Treat the current 200,000-token context window as a ceiling, not a target. Remove vague or overlapping skills before the shelf becomes difficult to inspect.
How do I choose a skill Claude Code will actually use? Choose one whose description names the task you repeat. Claude Code chooses from that description alone, so a vague description can leave a useful skill untouched.
Do it now
Run npx agentmods add <slug> to install one skill, then browse all Claude Code skills.