Every skill page on this site that has an install block shows a prompt you paste into the agent chat. That prompt is text. It does not, by itself, download a binary, open a port, or change your operating system. What happens next depends on the agent you pasted it into and on what the prompt tells that agent to do. Treat the paste as running instructions, not as bookmarking a page.

Before you paste

Read the prompt on the skill page all the way through. You are about to hand it to a system that can edit files in the project you have open. Confirm three things:

  • The prompt matches the job you opened the page for. If you landed here from search, re-read the title. Search pages are not reviews.
  • You know which project folder is open. A documentation skill pointed at the wrong repository will rewrite the wrong files.
  • You are willing to throw the chat away if the result is bad. Do the first run in a branch or a copy, not on the only copy of a client deliverable.

If the prompt mentions an API key, a cloud login, or a local export of personal chats, stop and read when not to install a skill. Pasting the prompt is the moment the risk becomes real.

Installing in a chat agent

Open a new chat so old instructions are not still in the context. Paste the prompt as the first message, then add one concrete task in your own words. For example, after a document skill's prompt, ask it to create a single empty outline with headings you specify. Do not start with your largest file.

Watch the first reply for two signals. The useful signal is that the agent repeats the constraint you care about, such as "I will not invent citations" or "I will only touch Markdown under docs/." The bad signal is that it immediately wants a credential, wants to run a destructive command, or claims it has installed software you did not see it install.

If the agent you use stores "skills" or "rules" as a named object, the prompt may tell it to save that object. In that case the text lives in the agent's settings, not only in the chat. Removing the chat later will not remove the skill. That distinction is the whole point of the uninstall section below.

Installing in an editor agent

Cursor and similar editors can apply a prompt to the current workspace. Paste it, then ask for a dry description of the files it would change. Only after that description matches your intent should you let it edit. Commit or stash your work first so a bad edit is a revert, not a reconstruction.

Some skills in Development assume a particular stack. A Spring Boot skill will give confused advice in a Python repository, and it may still sound confident. If the first answer mentions the wrong framework, the install "worked" and the skill is simply the wrong one. Remove it. Do not add a second skill to cancel the first.

How to tell the skill is active

Ask a question the skill should refuse or reshape. A legal retrieval skill should refuse to invent a statute. A quote-review skill should ask for the quote instead of pricing a renovation from imagination. A writing skill should follow the format it advertised. If the reply looks like the agent's ordinary behavior, the prompt was ignored or buried under a longer system prompt. Paste it again in a fresh chat before you conclude the skill is useful.

Also check the agent's own skill or rule list if it has one. The name on that list should match the card. If the agent saved a different name, you will not find it later when you want it gone.

Updating

Skills on this site show an updated date and sometimes a version when the upstream listing has one. An update is not pushed to your agent. The directory does not call your editor. To update, open the skill page, read the new prompt, and replace the old text in the agent. Diff the two prompts yourself if the job is important. Publishers change instructions. A prompt that used to stay inside one folder may now ask for network access.

If the upstream link on the card is dead, do not hunt for a similarly named repository and assume it is the same project. Send the page URL through the contact form and leave the old prompt in place until you know what replaced it.

Removing

Deletion has two layers.

The chat layer is easy. Close the conversation or start a new one without pasting the prompt. That is enough when you only pasted text into a single session.

The settings layer is the one people forget. If the agent saved a skill, a rule, a project instruction, or a file such as a rules document in the repository, delete that object too. Search the repository for a distinctive sentence from the prompt. If the skill wrote itself into the repo, that sentence will show up in git, and teammates will inherit it on the next pull. Remove the file in a commit whose message says you are dropping the skill.

After removal, ask the agent to do the same small task again. The reply should no longer follow the skill's format. If it does, a second copy of the prompt is still in a user-level rule or a team rules file. Keep looking. Two copies are a common reason "uninstall" seems to fail.

What this site does not do

We do not run the install prompt on your machine. The copy button only copies text. We do not store your keys. A skill page that tells you to paste a key is instructing you to give that key to your agent and, by extension, to whatever tool the agent calls. That is between you and the publisher. Our privacy policy covers this website, not the third-party API the skill wants to call.

Install one skill, prove it on a small example, and keep the uninstall path as short as the install path. A skill you cannot find in the agent's settings is a skill you should not have saved.

A concrete install and removal

Take a documentation skill aimed at a repository, the sort of thing filed under Knowledge Management. Before pasting, create a branch called skill-trial and add a single comment in a source file that a human would expect the docs to mention. Commit that comment so the trial has a parent you can reset to.

Paste the prompt in a fresh chat and ask only: "Update the docs for the comment I just added. Do not edit source files." The useful outcome is a documentation diff. The failed outcome is a rewritten source file, a new dependency, or a paragraph that describes a feature the comment does not mention. If you get the failed outcome, do not negotiate with the skill by adding "please really do not." Remove it.

Removal on this trial looks like this. Close the chat. Search the repository for a sentence that appears only in the prompt, including hidden folders the editor uses for rules. Search your user-level agent settings for the skill name on the card. Delete each hit. Check out the branch's parent commit if the skill already wrote files, or revert the doc diff if the doc diff was wrong. Open a new chat with no prompt and ask what the comment means. The answer should be ordinary and should not follow the skill's template. If the template remains, a rules file is still loaded from a parent directory or a team settings repo. Look one level up.

That half hour is the real install cost. A card that hides its prompt behind "just click copy" has the same cost, you simply have not paid it yet. The first week plan spends a day on exactly this loop so the rest of the week is not spent guessing which prompt is speaking.

Updating without losing the trial

When the card's updated date changes, download nothing automatically. Copy the new prompt into a diff against the text you saved. Read only the added lines. If a new line grants network access or broadens the folders the agent may edit, treat it as a new install and repeat the trial on a fresh branch. If the new lines are wording, replace the saved prompt and keep the skill. Store the prompt next to the branch name in your notes so the next update has something to diff against. The directory does not keep that history for you.