An agent skill, on this site, is a packaged prompt and a short description that you paste into Claude, Cursor, or a similar coding agent so it follows a narrower job. The Agent Skills directory holds thousands of them. That number is not a recommendation. Most of the cards are a name, a publisher, an install prompt, and a paragraph. Choosing well means ignoring almost all of them.

Start from the job, not the category

Write one sentence that a colleague could act on. "Turn a messy client brief into a one-page visit proposal" is a job. "Be better at business" is a mood. If you cannot write the sentence, you do not have a skill problem yet. You have an unclear task, and installing a skill will only add instructions on top of the confusion.

Then open the category that matches the sentence, not the category that sounds impressive. A proposal belongs in Business Operations. A Spring Boot endpoint belongs in Development. A lesson plan belongs in Education. A packet capture belongs in IT Ops & Security. The category essay at the top of each of those pages says who the section is for. Read that before you read the cards.

Read the card in this order

  1. The name and the publisher. A name that promises "pro", "ultimate", or "complete" is advertising. The publisher line tells you whether a company stands behind it or an anonymous account uploaded it yesterday.
  2. The install prompt. If the prompt is a wall of rules that contradict each other, the skill will fight the agent you already configured. Prefer a prompt you could explain out loud in under a minute.
  3. What the skill claims to touch. Documents in a cloud suite, your local repository, a browser, a chat export, or a paid API are different risks. The when not to install guide is the checklist for that sentence.
  4. The upstream link. This directory is an index. The source page is where the license and the real README live. If the link is dead, skip the card.

Downloads and the directory score are last. They tell you the listing has been copied around. They do not tell you the prompt is safe, current, or right for your repository. How ranking works explains the sort keys so you can stop treating them as a review.

Prefer a skill that names an artifact

Good skills end in a file you can check: a PDF, a diff, a table, a lesson plan, a test. Weak skills end in "insights" or "a better workflow." When two cards claim the same job, keep the one whose prompt says what the output looks like and what it refuses to invent.

Look at the editor's notes when they exist. Those notes are written here, and they say who should skip the item. A card without a note is still only a listing. Do not read a long publisher description as our review. The publisher wrote it to get the install.

One skill at a time

Install a single skill, run it on a real input you already understand, and compare the output to what you would have written yourself. If it saves time on that input, keep it. If it hallucinates a citation, touches files you did not mention, or asks for a key you do not want in the chat, remove it the same day. The install and remove guide is the short procedure.

Stacking five skills "just in case" makes the agent longer, slower, and harder to debug. When the answer is wrong you will not know which prompt caused it. The first week guide suggests a smaller starting set for people who write code.

A short list of rejects

Skip a skill when any of these are true:

  • The job sentence and the skill's examples do not overlap.
  • The prompt tells the agent to ignore other instructions, hide its steps, or keep going after an error.
  • The only way to try it is to paste a production credential into the chat.
  • The upstream page and the card disagree about the license or the publisher.
  • You wanted a model comparison or a plugin for DeepSeek Harness. Those are different shelves: models and plugins.

What "curated" means here

The directory is curated in the weak sense: entries came from public listings and were grouped into categories. It is not curated in the strong sense of "a person tried every prompt." We write full notes for a small set of listings and long guides for the decisions that repeat. If you want a correction on a card, use the contact form. If you want the policy on aggregated text, read the terms and about.

Choose the skill that matches one job, that you can remove, and that you can check against an example you already trust. Everything else in the directory can wait.

A worked pass through the directory

Suppose the sentence is "check a renovation quote for missing electrical items before I pay the deposit." That sentence names an artifact, a quote, and a failure, a missing trade. It belongs in Life Service, not in data analysis and not in a general writing skill. Open the category, ignore the download column for a moment, and look for a card whose examples mention budgets, unit prices, or trade scopes. A card about novels or about Spring Boot is a search mistake even if the word "review" appears in both.

Read the install prompt and mark every noun that is an input. If the prompt demands the full chat history of the household, you do not have to provide that to check one PDF. If the prompt demands a cloud key for a service you do not use, it is the wrong tool for a local PDF. The stop list is the formal version of that mark-up.

Then pick one card and invent a tiny fake quote: three line items, one of them missing a unit. Run only that. If the skill flags the missing unit, it earned a second run on the real quote with names removed. If it invents a city price list you did not give it, remove it. A confident price is worse than no skill, because you might pay the number.

Do the same pass for a programming job and the shelf changes. "Add pagination to this Flask list" is Development. The fake input is a route that returns every row. The success test is a diff that defaults the page size and does not rewrite the template's unrelated copy. Notice the pattern is identical and the category is not. People waste the afternoon in the wrong category because the top of AI Agent sounds more advanced than the category that names their file type.

What to do with two finalists

When two cards both match the sentence, prefer the one that lists a refusal. "I will not invent a statute," "I will not add packages you did not mention," "I will not send the file anywhere." A refusal is a test you can run. The other card, the one that only lists capabilities, will agree with whatever you ask, including a bad idea.

Prefer the card whose publisher you can find again next month. A company documentation skill with a stable source URL is easier to update than an anonymous upload with the same feature list. Prefer the shorter prompt when both are honest. Extra adjectives are instructions the model will half-follow.

Write the winner's name in your own notes with the sentence it serves. The directory will not remember that you chose it for quotes and not for contracts. Future you, staring at the same category, will otherwise install a second skill for the same job. The checklist is what you run on that second impulse.