Claude Code skill management. Topics — writer (create), lint (validate + fix frontmatter), merge (combine related), dedup (find duplicates), convert (agent → skill), architecture (multi-topic structure), upgrade (enhance + add topics), route (topic placement), trigger (declare + auto-register hooks), find (discover via npx skills CLI), graph (extract depends-on + topic body Skill calls into Edge Table + Mermaid + dispatched d3 force-directed render), language (per-skill language consistency enforcement), portability (public skill cross-ref + vendor isolation), publish-scope (published skill scope review before extending), invoke-discipline (slash command → Skill tool call, multi-topic Read, post-decision auto-invoke, interactive script, vendor dispatch). Use when: "skill writer", "skill lint", "skill merge", "skill dedup", "create skill", "frontmatter fix", "multi-topic skill", "convert agent", "skill upgrade", "add topic", "topic route", "trigger compile", "hook auto register", "find
Coding
Skills Manager CLI
Try itDrive the Skills Manager CLI (`skm`) to initialize the hub, adopt unmanaged skills, list/enable/disable skills per AI tool, and doctor/fix symlink sync. Use whenever the user or an agent needs to manage skills from a terminal, SSH session, CI job, or headless machine; when a skill is missing in Claude Code, Codex, Cursor, Gemini, OpenCode, or any other supported tool; when asked to run skm / Skills Manager CLI; or when setting up Skills Manager without the GUI. Do not use for `npx skills` / skills.sh search (different CLI), for writing SKILL.md content (skill-creator), or for marketplace browse/install (not in skm).
What it does
Drive the Skills Manager CLI (`skm`) to initialize the hub, adopt unmanaged skills, list/enable/disable skills per AI tool, and doctor/fix symlink sync. Use whenever the user or an agent needs to manage skills from a terminal, SSH session, CI job, or headless machine; when a skill is missing in Claude Code, Codex, Cursor, Gemini, OpenCode, or any other supported tool; when asked to run skm / Skills Manager CLI; or when setting up Skills Manager without the GUI. Do not use for `npx skills` / skills.sh search (different CLI), for writing SKILL.md content (skill-creator), or for marketplace browse/install (not in skm).
The skill document
Skills Manager CLI (skm)
skm is the terminal front-end of Skills Manager. It reads and writes the
same ~/.skills-manager/config.json and the same hub/tool symlinks as the
desktop app, so GUI and CLI are interchangeable.
It does not create, edit, search, translate, tag, or marketplace-install
skills. Those stay in the GUI (or you write files into the hub yourself, then
skm enable).
Preconditions
- Resolve the binary. Prefer
skmonPATH. If missing, say so and point at install (Settings → Command Line Tool, or a release archiveskm-.tar.gz/.zip). Do not fall back tonpx skills. - Confirm with
skm --version. A stale binary from an older app release should be updated from Settings or a matching release. - Every command except
initrequires an initialized config. If stderr says so, runskm init --jsononce — it is idempotent and will not clobber an existing config.skm initand Settings → Install CLI also copy this companion skill into the hub and enable it for every currently active tool, so you usually do not need to install it by hand.
Hub path is fixed at ~/.skills-manager/skills (%USERPROFILE%\.skills-manager\skills
on Windows). save() always rewrites skills_dir to that default; there is
no supported way to pick another hub from the CLI.
Agent conventions
These exist because skm is designed to be called by agents, not just humans.
- Prefer
--jsononinit,list,doctor,fix, andadopt(apply). Data is one JSON document on stdout. Errors go to stderr as{"error":"..."}when the command was in JSON mode, otherwiseerror: .... A command can print a result on stdout and still exit 1 (fixwith failed repairs) — parse stdout, then trust the exit code. enable/disablehave no--json. Success is a one-lineenabled '' for ''(ordisabled). Failure iserror: ...on stderr, exit 1.- Never wait on a TTY prompt.
adoptwithout--yesand without--jsonblocks on[y/N]. Pass--yeswhen applying.--jsonalso skips the prompt (and therefore applies) — still pass--yesso the intent is obvious in the command line. skm adopt --json --dry-runis not a machine-readable preview: if candidates exist it prints nothing. Preview withskm adopt --dry-run(human text), then apply withskm adopt --yes --json.doctor --jsonexits 0 even when there are issues. Gate on.issues_count, not the exit code.fixwithout--yesdoes not repair. JSON then looks like{"applied": false, "issues_found": N, ...}. Apply withskm fix --yes --json.- Write commands (
init,adopt,enable,disable,fix --yes) take an advisory lock onconfig.jsonand fail fast if anotherskmis mid-write. Retry once. The desktop app does not take this lock. - Do not enable a skill for every detected tool unless the user asked. Pick the tool they named, or the host agent (see Which tool).
JSON shapes, exit codes, and the dry-run/apply matrix live in
references/json.md. Read that before parsing output or writing a script.
Command map
| Intent | Command |
|---|---|
| First-run / is it initialized? | skm init --json |
| What skills exist, and where are they linked? | skm list --json |
| Same, one tool only | skm list --tool --json |
| Turn a hub skill on for a tool | skm enable --for |
| Turn it off (remove the link, keep the hub copy) | skm disable --for |
| Which tools are installed, any broken links? | skm doctor --json |
| Repair what doctor reported | skm fix --yes --json |
| Preview leftover real dirs in tool skill folders | skm adopt --dry-run |
| Move those dirs into the hub and relink | skm adopt --yes --json |
and accept an exact id, an exact instance_id, or a
unique prefix (claude → claude-code). Ambiguous prefixes error and
list the matches — do not guess, rerun with a longer token. Exact tool id
wins before prefix (trae is Trae, not Trae CN). Full ids and collision
notes: references/tools.md.
Workflows
First-run on a machine with no GUI
command -v skm
skm init --json
skm adopt --dry-run
skm adopt --yes --json
skm list --json
skm doctor --json
init on an already-initialized config returns
{"already_initialized": true, ...} and does nothing else. If
config.json exists but cannot be parsed, init refuses to overwrite it
— move or fix the file, then retry.
Make an existing hub skill available to a tool
skm list --json # find id / instance_id
skm enable --for
skm list --tool --json # confirm enabled_for
Enabled state is derived from the link on disk, not stored as a separate
flag. enable creates the symlink (Windows: symlink → junction → tracked
copy). disable removes that link; the hub copy stays.
If the same skill id exists in both global and the active project, a bare
id is ambiguous. Use the instance_id from list: global: or
project::.
Add a brand-new skill (CLI cannot scaffold it)
- Write
~/.skills-manager/skills//SKILL.mdwith YAMLname+description. A directory is a skill if it containsSKILL.md,skill.md, ormeta.json. skm enable --forskm list --tool --jsonto confirm.
Do not copy the folder into ~/.claude/skills (or any other tool dir)
yourself — that bypasses the hub and adopt will later see it as an
unmanaged real directory.
Adopt skills that already live in a tool directory
adopt looks at active tools (enabled + detected) for real
directories that are not symlinks/junctions and not hidden. It moves
each one into the hub, then puts a link back at the original path.
- Duplicate hub names are skipped and left in place (contents are not merged — they may differ).
- Preview first (
--dry-run), then apply (--yes --json). - After apply,
listshould show the new ids as enabled for the source tool.
Diagnose "the skill is not showing up in Claude / Codex / …"
skm doctor --json
skm list --tool --json
Read the results in this order:
- Binary missing or not on
PATH→ install, reopen the shell. config_initializedfalse →skm init --json.- Tool
detected: false→ the tool's config dir is not on this machine. CLI does not re-detect existing tools on every call; a tool installed afterinitmay still shownot installeduntil the GUI redetects or the config is updated. Do not try to invent a custom tool fromskm. - Tool
enabled: false→ CLI cannot flip that flag. Use the GUI Tools page, or say so. - Skill absent from
list→ it is not in the hub (and not in the active project). Write it into the hub oradopt. - Skill present but not in
enabled_for→skm enable --for. link_issues> 0 orissues_count> 0 →skm fix --yes --json. Iffailed_count> 0, report thefailed[].messagevalues; do not loopfix.- A wrong-target or deleted link is treated as disabled, not as a
doctor issue. Re-
enableto recreate the link over the tampered path.
fix only repairs what should_report_sync_issue flags. It will not
delete a real (non-link) directory sitting in a tool's skills folder —
that is what adopt is for.
Which tool
If the user named a tool, resolve it via references/tools.md and pass
--for . If they did not:
| Host | --for |
|---|---|
| Claude Code | claude-code |
| Codex | codex |
| Cursor | cursor |
| Gemini CLI | gemini |
| OpenCode | opencode |
| Unsure | skm doctor --json, pick detected && enabled |
Prefix shortcuts that are unique today: claude → claude-code. Do not
use trae when the user said Trae CN (trae-cn), or qoder when they
said QoderWork CN (qoderwork-cn).
Out of scope (do not fake these with skm)
| Need | What to do instead |
|---|---|
| Search / install from skills.sh | npx skills / find-skills, or the GUI Marketplace |
| Author SKILL.md | Write the file; this skill only covers linking it |
| Tags, notes, favorites, translation | GUI |
| Add/remove a custom tool, bind a project | GUI |
| Enable or disable a tool itself | GUI Tools page |
| Change the hub directory | Not supported; always ~/.skills-manager/skills |
| Marketplace publish / ClawHub | GUI |
Install locations (when skm is missing)
The desktop app copies a bundled binary:
- macOS / Linux:
/usr/local/bin/skmif writable, else the first of~/bin,~/.local/binthat is already onPATH, else~/.local/bin/skm(then the folder may need to be added toPATH) - Windows: next to the app executable (
skm.exe)
Standalone: GitHub Releases skm-.tar.gz (.zip on Windows),
extract, put on PATH.
Related skills
Toggle skills across major coding agents
Audit and converge local skills across Codex, Claude, OpenClaw, OpenCode, workspace skills, and shared agent libraries; compute a hygiene score, classify sha...
Sync one skill across eight publishing channels
Install skills from multiple skill hubs with one tool: clawhub.ai (default), skillhub.cn (SkillHub, China-optimized), skills.sh (npx skills, GitHub-based), t...
Expose installed agent Skills as HTTP(S) REST API services. Runs a persistent FastAPI server that auto-generates an API endpoint per Skill, with sync/async execution, webhook callbacks, and a multi-engine sub-agent executor (OpenClaw / Claude Code / Codex CLI / LLM fallback). Ships a bilingual (EN/ZH) web management console. HTTP by default (zero-friction), optional HTTPS with self-signed SAN certificates for production. Use when you need to serve Skills over HTTP, call Skills remotely, or expose Skill capabilities to external systems.