Skill Libraries give you a personal, reusable collection of instructions that LatchLoop agents can follow across tasks. You can manage the content in LatchLoop or connect a GitHub repository as its source.
This guide explains how to create, connect, use, update, and safely run skills.
A Skill Library is a user-owned collection of standard Agent Skills. The library is mutable, which means you can add, update, and remove its skills over time.
Each skill is a folder with a required SKILL.md file. It can also include:
A skill is a reusable procedure or set of instructions. It is not the report, plan, code change, or other output produced when an agent follows those instructions.
If a library is available to All projects, its skills can be discovered from every project you are permitted to use. LatchLoop does not load every skill’s full instructions into every prompt. It shows compact details first and reads a skill only when it is mentioned or needed.
Skill Libraries in this release are personal and user-owned. They are not public or team-owned.
When you choose All projects or Selected projects, you control where you and your agent can discover and use the library. This setting does not grant another person access to your library.
Keep these access rules in mind:
The library list shows the information you need to understand its current condition, including:
A managed library stores the skill content you accept in LatchLoop and gives you version history, editing, restore, and export controls.
You can change project availability later without changing the skills themselves.
Choose the method that matches what you have:
SKILL.md for a skill that has no additional resources.LatchLoop validates the skill names, required files, and SKILL.md frontmatter—the identifying details at the top of the file—before import. If an upload contains both valid and invalid skills, the confirmation explains that only valid skills can be imported. Review the warnings before continuing; invalid skills are not added.
Warning — replacement is destructive: Before choosing Replace entire library, export the current library or confirm that its version history contains what you may need. Carefully review the list of skills that will be removed before confirming.
When an imported skill has the same name as an existing skill:

You can edit the skill’s instructions and text resources directly in LatchLoop.
SKILL.md or a text resource in the editor.Binary resources, such as many images or packaged files, cannot be edited in the text editor. Download a binary resource to inspect or update it with an appropriate app, then add the updated resource to the skill.
To remove a skill:
Warning — deletion removes the skill from future use: Export the library first if you may need a separate backup. For a managed library, you can also restore an earlier library version.
Every successful managed-library change creates a new immutable version. Earlier versions remain unchanged, giving you a reliable history even as the current library evolves.
Restoring does not rewrite the old history. It creates a new current version from the earlier accepted content.
To keep a portable copy, select Export ZIP. The export contains the accepted skill content from the selected managed-library version.
Warning — restoring changes the current library: Review the diff and affected skills before confirming. Export the current version first if you want an additional backup.
Before connecting a repository, install and connect the LatchLoop GitHub App with access to that repository. See GitHub Integration for setup help.
To connect the library:
GitHub remains the source of truth. LatchLoop indexes and caches valid skills so the agent can discover and use them, but it does not turn the connected library into a separately managed copy.
Under either policy, an agent can only propose a change. You must explicitly approve it before LatchLoop writes to GitHub.
Warning — direct commits bypass pull-request review: Use Direct commit only when writing straight to the selected branch is appropriate. Protected-branch rules and repository permissions still apply.
The status shows whether the indexed skills reflect the connected source:
A reachable repository can appear as an empty library when the selected branch and root directory contain no valid skills.
From the library details, you can:
Warning — deletion removes the library connection from LatchLoop: Confirm that no current tasks depend on it. Deleting the library connection does not delete the GitHub repository, but its skills will no longer be available through this library.
Do not edit GitHub-backed skill files in the managed-library editor. Change them in GitHub, or approve an agent-proposed GitHub change, and then sync the library.
@ and begin entering the skill’s name.Plugin skills and your personal skills can appear in the same search. LatchLoop initially gives the agent compact information about available skills. It loads the full SKILL.md, or a particular resource, only when you mention it or the agent needs it. This keeps prompts focused and reduces conflicts between unrelated instructions.
For example:
“Follow @Release checklist when preparing this change for review.”
“Use @Customer interview analysis and read its
references/scoring-guide.mdresource before summarizing these notes.”
If a library is available to All projects, this discovery works across every project you can use. It still does not preload all of the library’s instructions into each task.

An agent can propose creating, updating, or deleting a skill, or adding a resource. It cannot silently save the change to a managed library or GitHub-backed library.
To review a proposal:
Approval uses the library state the proposal was created for. If the managed library or connected GitHub branch changed afterward, the proposal becomes stale and fails safely instead of overwriting newer work. Ask the agent to create a new proposal against the latest version.
Warning — review deletions and GitHub writes carefully: Approval can remove skill content or write accepted changes according to the library’s GitHub write policy.

Scripts are never run merely because a library is imported, indexed, mentioned, or read. Cloud script execution uses a dedicated script-execution tool and requires a fresh, one-time approval for every invocation.
When the agent requests a script run:
Warning — a script executes instructions: Approve only when you recognize the skill, script path, runtime, and arguments and want that exact invocation to run. A later invocation requires a new approval.
Approved scripts run in a restricted cloud sandbox. They do not inherit your secrets and do not have network access. Local or desktop skill-script execution is not supported in this release.
A skill’s allowed-tools information cannot bypass LatchLoop’s approval requirement or other safety controls.
Managed libraries are portable: export accepted content as a ZIP, and use version history to restore an earlier accepted version.
LatchLoop checks imports before accepting them. Validation includes skill names and SKILL.md information. Unsafe or unsupported input is rejected, including:
Files outside valid skill folders are not indexed or included in exports. This keeps the library focused on accepted skill content rather than unrelated files that happened to be in an uploaded folder or repository.
SKILL.md file.SKILL.md.SKILL.md is invalidSKILL.md and belongs to a valid skill folder.@ searchThis release does not yet support:
There is also no public sharing link for a Skill Library. Export a managed library as a ZIP when you need a portable copy, and share it only through a channel appropriate for its contents.