Reuse procedures with skills

Skills are local SKILL.md files that give the agent a named procedure, description, and optional invocation controls. Use them once with /name, pin them to a session, or let the agent load one when its description matches the task.

Discovery boundary: a session reads personal skills (~/.claude/skills, ~/.claude/commands), then the project's (.claude/skills, .agents/skills, .claude/commands) in the working directory and every parent up to the repository root. A chat in the Default project reads the personal ones. Your skills are Claude Code's own files: Localix writes ~/.claude/skills and a project's .claude/skills, and Claude Code loads what it writes.

What you can do

Task	Mechanism	Boundary
Run a skill once	Enter `/name [arguments]` or select it from the web slash menu.	Selection inserts a one-shot command; it does not pin the skill.
Let the agent run it	Write `/name` as its own word later in a message (`go ahead and /deploy to staging`).	Permission for that message, a user-only skill included; the agent decides whether to.
Keep a skill active	Pin it to the session through the session-skills API.	The full skill body is added to that session's runtime prompt.
Create or revise a skill	Use Add skill on Store → Library → Yours (or Project), then edit it on its page.	Plugin-bundled skills are read-only; they follow their plugin.
Let the agent load a skill	The runtime lists names and descriptions; the agent can fetch a matching body with the `Skill` tool.	Bodies are loaded on demand rather than placing every skill in the prompt.
Archive stale skills	Press Archive unused on Yours or Project, check the preview, confirm; or Archive one skill.	Archived directories are excluded from discovery until restored.
Deny one agent a skill	Switch it off on the agent's Store page, Access tab.	The switch writes one `skill:<name>` rule; the file and every other agent are untouched.
Control availability	Switch the skill on or off in Store → Installed, or per agent on its Access tab.	Disabling removes access by name, including pins and keyword triggers; files remain.

How it works

Terms

Term	Meaning
Filesystem skill	A directory containing a readable `SKILL.md`.
Project skill	A skill under `<working-directory>/.claude/skills`.
Personal skill	A skill under `~/.claude/skills`, used by scratch chats.
Plugin skill	A read-only skill contributed by an enabled plugin.
Session pin	A skill name whose full body is injected for one session.
Curator pin	Protection from automatic stale-skill archiving.

Discovery and precedence

Session scope	Folders read, in precedence order	Additional source
Project	`~/.claude/skills`, `~/.claude/commands`, then in the working directory and each parent up to the repository root: `.claude/skills`, `.agents/skills`, `.claude/commands`	Enabled plugin skills
Default project	`~/.claude/skills`, `~/.claude/commands`	Enabled plugin skills

A command file .claude/commands/<path>.md is a skill named by its path with each / a : (commands/git/commit.md is /git:commit); it takes a skill's frontmatter except name and paths.

The scanner walks nested directories to find SKILL.md, up to six levels and 2,000 entries per root. Project symlinks must resolve inside the session's working directory. Personal skill symlinks may point outside the home directory; plugin skill links stay inside their declared root.

Built-in slash commands keep precedence over a skill with the same name. As in Claude Code, a personal skill takes precedence over a project skill of the same name, a nearer project folder over a parent's, a skill over a command file, and within one folder Localix's own skills over .claude/skills and portable .agents/skills. Two skills of one name in the same place keep the first sorted one, with a duplicate-name warning. Portable project skills are read sources: the Localix manager does not delete or archive them. Saving the same name creates a native Localix copy that shadows the portable procedure.

Loading and invocation

The runtime initially gives the model a bounded roster of skill names and descriptions. The roster lists each description up to 1,536 characters, as Claude Code does, within an 8,000-code-unit budget: descriptions shorten to bare names before any name is omitted with a notice. The Skill tool can load one body or save a skill; plugin skills can be read but not saved through the tool. Listing and removing skills belong to the Store. Default delegated workers can also retrieve skills. An agent's explicit tool list and capability denials still control availability; pinning a skill supplies its body even without the retrieval tool.

One-shot /name invocation expands supported arguments before submitting the prompt. Filesystem skills can also opt into bounded local command expansion and confined file-reference expansion. A skill with context: fork runs in a child session, typed or loaded by the agent, as the subagent type its agent: names (Explore, Plan, a plugin agent, or an agent file from .claude/agents or ~/.claude/agents; the invoking agent by default). Other skills run in the current session. Child sessions cannot fork skills into more child sessions.

Session pins persist with the session record. The runtime injects each resolvable pinned body under an active-skills prompt section on subsequent turns, within a shared 32,000-code-unit budget. Unavailable or unreadable pins are omitted. Agent-profile preloads use the same prompt section.

A skill can declare keywords to load its body when a phrase matches the latest non-internal user message. Matching is case-insensitive, respects word boundaries, and tolerates whitespace within phrases. These triggers do not inspect tool output. Hiding a skill from the model roster does not disable its keyword trigger; agent availability rules do.

Slash invocation resolves filesystem skills and enabled plugins' skills (/<plugin>:<skill>, or the bare name when only one plugin has it); a plugin's command files are skills too. The model-facing Skill tool and agent preloads reach the same skills; session pins resolve filesystem skills.

The supported frontmatter, substitution tokens, tool policy, and command-expansion limits belong to the skill authoring guide.

Management and curation

The Store creates, edits, removes, archives, restores and curator-pins the skills in ~/.claude/skills and the open project's .claude/skills, the folders Claude Code reads too: Add skill on Library → Yours (yours) or Library → Project (the open project's), and the skill's own page for the rest. Its Settings tab holds the description and the instructions; Pin, Archive and Restore are buttons beside them. An edit replaces the managed description and body and preserves other frontmatter. A skill's name is its identity and is chosen when it is added.

A skill that is stale or close to it says so at the top of its page: "Not used for 95 days. Archived on 15 Jun 2026 unless pinned." with Pin beside it (a project's skill says to use Archive unused instead, since nothing archives it on its own). The page warns within two weeks of the pass that archives it.

Skills from the portable .agents/skills, from command files and from a parent folder's .claude/skills are listed read-only, and say where they live; a plugin's skills follow the plugin. A .claude/commands file is discovered as a skill, so there is no separate commands kind. The agent's Store page (Access tab) keeps the switches.

.claude/skills is shared with Claude Code, whose use of a skill Localix cannot see. So the scheduled pass and Archive unused consider only skills created in Localix (the Store's Add skill, the agent's Skill tool, or skills moved from the old .localix/skills). A skill Claude Code or you created by hand there never shows an "archived on" notice and is never archived by itself; it can still be pinned, archived from its page and restored.

The curator classifies an unpinned skill created in Localix as stale when it has never been used for more than 30 days or was last used more than 90 days ago. Automatic checks run once the app is idle after startup and then daily, but an archive pass runs at most once every seven days. The first automatic observation only saves the schedule marker and waits a full interval before archiving. When a pass archives skills, one Event Center row names them (one skill: "Archived the review skill"; more: "Archived 3 unused skills") with Open to restore, which opens that skill's page (or the skills shelf) where Restore is.

Archive unused on Yours and on Project is the same rule on demand. It first shows a preview, each skill that would be archived with why ("Last used 100 days ago", "Never used, 45 days old"), and archives nothing until you confirm. Confirming archives the previewed skills that are still stale: one you pinned or used meanwhile is left alone. Pinned skills never appear in the preview.

One-shot slash preprocessing and successful skill.get record usage on a best-effort basis; they do not prove that the procedure completed. Pinning and keyword injection do not update this usage counter. A never-used skill's age comes from its usage record or file modification time.

Curator inventory covers the .claude/skills folder of the selected scope, without the runtime's personal fallback or portable sources. Archiving moves the directory to ~/.localix/skills-archive/user/<name> (a project's: .../<projectId>/<name>), outside the folders Claude Code reads, so an archived skill never loads there; restoring moves it back (and refuses when a folder of that name exists) and records a use. Archiving can replace an archived skill of the same name.

Curator pins protect automatic passes, but manual archive remains available.

Configuration

Setting or surface	Effect
`<working-directory>/.claude/skills`	Stores skills for that project scope.
`~/.claude/skills`	Stores skills for scratch chats.
Agent Store page, Access tab	Switches skills per agent.
Web slash menu	Inserts a one-shot `/name` invocation.
Session-skills API	Replaces the session's persisted pinned-skill list.
Agent skill rules	Enables or disables access to a named skill for that agent.

Limits

Personal and project skills are merged, personal first, as in Claude Code: a personal skill shadows a project skill of the same name.
Personal .agents/skills is not read. Skills in a folder below the working directory load once the session works with a file there, as in Claude Code.
Skills read from .agents/skills, .claude/commands and a parent folder's .claude/skills are read sources: the Store writes, deletes and archives only the open scope's .claude/skills.
Project and plugin symlinks cannot extend discovery outside their safety root; personal skill links can. Archived and hidden subdirectories are excluded, and a discovered skill's assets are not scanned for nested skills.
Localix supports a defined SKILL.md contract, not every convention used by other tools. Check the authoring guide before relying on a field or expansion.
The web slash menu runs a skill once and the web pinned-skill row can remove existing pins; adding a persistent pin is currently exposed only through the session-skills API.
A skill cannot be renamed in the Store: add the new name and remove the old skill.
Filesystem skills may execute declared local shell expansions or read confined file references during one-shot invocation. Review skill files as executable local instructions.
Validator warnings are findings, not load blockers. Discovery rejects unreadable files, invalid UTF-8, files over 1 MiB, and empty Markdown bodies, including description-only files.
Automatic curator passes operate on personal skills. Use Archive unused on the Project shelf for a project's skills.
Curator pins prevent automatic archiving; they do not pin a skill into a chat session.
Model skill.get returns at most 64,000 bytes of body text, with a truncation notice and source path for longer procedures. This is separate from the 1 MiB file acceptance limit.