manage-skills git:20260830.50da461A

Immutable. This exact content is served forever at /api/v1/blob/3e93398da4af5b14.

---
name: manage-skills
description: Manage the user's shared agent-skill library via skills-manager-cli — install, update, remove, deploy or undeploy skills per agent, manage presets, organize tags, search, and adopt existing skills. Use this whenever the user wants Claude Code, Codex, Cursor, or another agent to gain or lose a skill, wants to organize the central library, or asks what is installed or deployed. Prefer this over direct agent-folder installs because Skills Manager preserves source metadata, preset membership, updates, and cross-agent deployment state.
---

## Before doing anything

1. **Resolve the CLI first, then use the path it prints.** Run this once (POSIX
   shell):

   ```bash
   D="$HOME/.skills-manager/bin"
   B="$D/skills-manager-cli"; [ -e "$B" ] || B="$B.exe"   # .exe on Windows
   if [ -s "$D/.version" ] && [ -x "$B" ]; then
     echo "$B"
   elif [ -s "$D/.version" ] || [ -e "$B" ]; then
     echo BRIDGE_BROKEN
   else
     P="$(command -v skills-manager-cli 2>/dev/null || true)"
     [ -x "$P" ] && echo "$P"
   fi
   ```

   **Substitute the printed path into every command below**, wherever the
   examples write `$SM`. Do not carry `$SM` as a shell variable: each command
   you run is a new shell, so an assignment made here is gone by the next one.

   The three outcomes:

   - **A path under `~/.skills-manager/bin`** — the desktop app published this
     copy, and the `.version` stamp appears only after it has been verified, so
     it always matches the app the user is running. Use it.
   - **`BRIDGE_BROKEN`** — something the app left behind is here but does not
     add up: an unstamped binary, or a stamp with no binary beside it. Either is
     what a copy that failed half-way leaves. **Stop.** Do not go looking
     for another CLI: that binary may predate a safety fix, and the machine has
     a desktop app whose version nothing here can match. Ask the user to open
     the Skills Manager app once, which republishes it.
   - **A path from PATH** — nothing was ever published here, so there is no
     stale copy to worry about: this is a CLI-only machine (a server install, a
     standalone download, a hand-built binary). Use it, but note it can be
     older than a desktop app if one is also installed.

   If nothing is printed at all, this skill doesn't apply — fall back to
   find-skills, or tell the user to install Skills Manager.
2. **Always pass `--json` when you parse output yourself.** Pretty-printed output is for the user; JSON is for you. Errors include `ok=false`, a stable `code`, and `message` on stderr with a non-zero exit code.

```bash
"$SM" --json skills list
```

### When a deployment is refused

A deploy that would overwrite something that is not ours is refused outright —
nothing at those paths is deleted, and nothing else in the batch is applied.
That failure is machine-readable, so report the actual paths rather than the
sentence:

```json
{"ok": false, "code": "TARGET_CONFLICT", "kind": "target_conflict",
 "message": "Refusing to deploy: 1 of 2 target(s) …",
 "details": {"conflicts": [{"path": "/Users/me/.claude/skills/db",
                            "reason": "is not a managed deployment"}]}}
```

Tell the user which path is in the way, that its contents are untouched, and
offer the two ways out: adopt it into the library (`skills adopt`), or move it
aside and retry. Never delete it for them.

## Mental model

There's **one central library** at `~/.skills-manager/skills/` that all agents share. Each skill has source metadata, preset membership, tags, and zero or more real deployments in agent directories. A **preset** is a reusable group; several presets may be deployed at the same time.

Keep these three states separate:
- **Library**: install/remove controls whether Skills Manager owns the skill.
- **Preset membership**: `presets add-skill/remove-skill` organizes the library only.
- **Deployment**: `skills deploy/undeploy` and `presets deploy/undeploy` control what an agent can actually see.

Internally, presets are still stored as scenarios for backward-compatible Git Backup. The CLI and UI call them presets.

## Install

```bash
# From skills.sh marketplace
"$SM" skills install vercel-labs/agent-skills@react-best-practices

# Any git URL (use /tree/branch/subpath form when the skill lives in a sub-directory)
"$SM" skills install https://github.com/anthropics/skills.git
"$SM" skills install https://github.com/foo/bar/tree/main/skills/baz

# Local folder
"$SM" skills install ./my-skill

# Force a source type when the ref is ambiguous
"$SM" skills install foo/bar --skillssh
"$SM" skills install ./looks-like/owner-repo --local
```

**Default is library-only** — the skill enters the DB but doesn't appear in any agent yet. Prefer an explicit follow-up deployment so scope is unambiguous:

```bash
"$SM" skills deploy <skill> --agent claude_code --agent codex
```

`--sync` and `--sync-preset` remain legacy shortcuts for the exclusive active-preset workflow.

**Ref resolution** is deterministic, no path-existence guessing:
1. Starts with `./`, `../`, `/`, or `~/` → local path
2. Contains `://`, ends in `.git`, or starts with `git@` → git URL
3. Matches `owner/repo`, `owner/repo/skill`, or `owner/repo@skill` → skillssh
4. Otherwise → error; pass `--local` / `--git` / `--skillssh` to disambiguate

**Always verify after install** with `skills list` or `skills show <name>` so you can confirm the skill landed and report the preset / sync state back to the user.

## Search

```bash
"$SM" --json skills search "react performance" --limit 5
```

Each result has `install_ref` (paste straight into `skills install`), `installs` (popularity proxy), and `skills_sh_url`. Show the top 1–3 with install counts before installing — anything with 10K+ installs is battle-tested; anything under 100 needs a careful look at the source repo.

## Update / Check

```bash
# Re-fetch one skill (git/skillssh re-clones, local/import re-imports source dir)
"$SM" skills update <skill-name-or-id>

# Re-fetch all eligible skills
"$SM" skills update --all

# Just probe remote revisions, don't touch files
"$SM" skills check --all
```

`check` is the dry-run partner of `update`. Local-only skills (no git source) are reported as `skipped: true`.

**An update replaces the skill's directory wholesale**, so anything written inside it that the new version does not have would be destroyed. When the CLI detects that, it applies nothing and reports the paths instead:

```jsonc
{ "name": "ppt-master", "refreshed": false,
  "held_back_removals": ["library: templates/mine.pptx"] }
```

The field is omitted entirely when nothing is held back, so test for its presence rather than for an empty array. `refreshed: false` *with* `held_back_removals` is **not a failure and not something to retry** — the skill is untouched and still on its old version. Show the user the listed paths and ask. There is no CLI flag to override this; only the desktop app can confirm and proceed, because only a person can say those files are expendable. The paths are prefixed with where they live (`library`, or an agent key for a deployed copy).

Note what this does *not* cover: a file the user edited that the new version also ships is reported as surviving, because its path survives — the update overwrites their edits silently. Warn anyone keeping local modifications inside a skill folder.

## Remove

```bash
# Always preview first when removing more than one
"$SM" skills remove <skill> --dry-run

# --yes is required for the actual delete; --json mode does NOT auto-confirm
"$SM" skills remove <skill> --yes
```

Remove deletes the central-library copy, all synced targets across agents, and the DB row. It's not reversible without re-installing.

## Deploy / Undeploy

```bash
"$SM" skills deploy <skill> --agent claude_code
"$SM" skills undeploy <skill> --agent codex
"$SM" skills deploy <skill-a> <skill-b> --agent codex --dry-run
"$SM" skills deploy <skill> --agent claude_code --agent codex
"$SM" --json skills status <skill>
```

These commands change real managed deployments without deleting the central-library copy or changing preset membership. `skills enable/disable` are deprecated compatibility commands and do not change deployment; never use them.

`skills deploy` and `skills undeploy` always require at least one explicit `--agent`, whether the command names one skill or several. `skills status` also reports target rows left by a custom agent that is no longer registered, so stale deployments stay visible and can be cleaned with an explicit undeploy while the row exists.

## Legacy exclusive sync

```bash
# Sync current active preset to all enabled agents
"$SM" skills sync

# Preview the target list — safe, no writes
"$SM" skills sync --dry-run

# Switch the one legacy active preset, then sync
"$SM" skills sync --preset "Web Dev"

# Only sync to a single agent (useful when one agent's directory got out of sync)
"$SM" skills sync --tool claude_code
```

## Adopt skills installed elsewhere

When skills already live in an agent's directory (e.g. installed via `npx skills add` or manual `git clone`) but aren't in the central library, pull them in:

```bash
# Dry-run scan first — lists candidates without writing
"$SM" skills adopt ~/.claude/skills --dry-run

# Adopt everything found — each becomes source_type=local (can't auto-update from git)
"$SM" skills adopt ~/.claude/skills

# Adopt a single skill and pin it to a git source so `update` works later
"$SM" skills adopt ~/.claude/skills/react-best-practices \
  --git-url https://github.com/vercel-labs/agent-skills/tree/main/react-best-practices

# Or pass --git-subpath explicitly when the URL is just the repo root
"$SM" skills adopt ~/.claude/skills/react-best-practices \
  --git-url https://github.com/vercel-labs/agent-skills \
  --git-subpath react-best-practices

# Skill lives at the repo root? Pass an empty subpath
"$SM" skills adopt ~/.claude/skills/my-skill \
  --git-url https://github.com/me/my-skill --git-subpath ""
```

`adopt` auto-excludes anything already in the DB or already a sync target, so it's safe to re-run. `--git-url` requires either a URL with a subpath (`/tree/branch/path`) or an explicit `--git-subpath` — without that, future `update` would re-clone the wrong directory, so the CLI refuses to guess.

`--git-url` only applies at the moment of adoption, while the directory is still unmanaged. Once a skill is in the library, use `set-source` below.

## Re-point a skill at a git source

```bash
# Preview: resolves the source and reports whether content differs. It clones to
# a temp dir, but writes nothing to the library or the DB.
"$SM" --json skills set-source <skill> --git-url you/skills --subpath my-skill --dry-run

# A GitHub /tree/ URL carries the branch and subpath already
"$SM" skills set-source <skill> --git-url https://github.com/you/skills/tree/main/my-skill
```

This is how a `local` skill becomes git-backed so `update` works, and how a
skill pointed at the wrong repo gets corrected. It updates the row **in place**,
so the skill id survives and the tags, preset membership and per-agent
deployments keyed to it all stay intact.

- The flag is `--subpath` here, not `--git-subpath` — that one belongs to `adopt`. Pass `--subpath ""` when the skill is at the repo root, which must itself hold a `SKILL.md`.
- `--branch` overrides a branch encoded in the URL.
- The report carries `content_changed` — a single boolean, **not** a file list. It compares the new source against the hash currently recorded for the library copy, not a fresh hash of the directory on disk, so edits made inside the central copy afterwards do not register as a difference. When it is `false` the library copy is left untouched and those edits survive; copy-mode deployments are re-synced either way.

**`--force` is destructive, and nothing stands between it and the user's files.**
A content difference is refused without it. With it, the whole skill directory is
replaced — staged, swapped in, and the old copy deleted — so anything in the
library copy that the new source does not ship is gone. Unlike `skills update`,
this path has **no** `held_back_removals` check: nothing is withheld, and nothing
asks. `--dry-run` cannot tell you which files are at stake, only that something
differs. Never pass `--force` on the user's behalf — report `content_changed:
true`, say that proceeding overwrites the library copy wholesale, and let them
decide.

## Tag

```bash
"$SM" skills tag add <skill> web frontend
"$SM" skills tag remove <skill> frontend
"$SM" skills tag set <skill> web frontend
"$SM" skills tag rename frontend web
"$SM" skills tag delete obsolete --dry-run
"$SM" skills tag delete obsolete --yes
"$SM" skills tag list <skill>   # tags on one skill
"$SM" skills tag list           # all distinct tags
```

Useful organization queries:

```bash
"$SM" --json skills list --untagged
"$SM" --json skills list --no-preset
"$SM" --json skills list --tag frontend
"$SM" --json skills list --preset "Web Dev"
"$SM" --json skills list --deployed-to codex
```

## Presets

```bash
"$SM" presets list
"$SM" presets current
"$SM" presets show "Web Dev"
"$SM" presets create "Web Dev" --description "Frontend work"
"$SM" presets update "Web Dev" --name "Frontend"
"$SM" presets delete "Old" --dry-run
"$SM" presets delete "Old" --yes

"$SM" presets add-skill <preset> <skill>...
"$SM" presets remove-skill <preset> <skill>...

"$SM" presets deploy <preset>                  # all enabled coding agents
"$SM" presets deploy <preset> --agent codex
"$SM" presets undeploy <preset> --agent claude_code
"$SM" presets undeploy <preset>                # every agent with target rows for this preset
"$SM" --json presets status <preset>
```

`deploy/undeploy` are additive and match the app's Preset pills. Explicit `presets apply/deactivate` commands remain for the legacy exclusive active-preset model; do not use them for normal "turn this preset on/off" requests.

The no-`--agent` defaults intentionally differ: deploy targets all installed, enabled coding agents; undeploy discovers the preset's actual target rows and removes them even when an agent is now disabled, uninstalled, or no longer registered. Use the no-agent undeploy for "turn this preset off everywhere."

Preset create/update/delete and add-skill/remove-skill are organization-only CLI operations. They never deploy or undeploy agent files implicitly.

## Health check

When sync misbehaves or a command errors in a confusing way:

```bash
"$SM" --json repo status   # base dir, skill / preset counts, active preset
"$SM" --json agents list  # detected agents and their target paths
"$SM" agents enable codex
"$SM" agents disable claude_code
```

`repo status` and `agents list` are read-only and are the first checks for "why isn't this skill showing up in Cursor" questions. `agents disable` is a real mutation: it removes every managed deployment for that agent. `agents enable` makes the agent globally available again and re-syncs the legacy active preset, if one exists; use explicit skill or preset deployment afterward when the requested state is additive.

Use `agents disable <agent>` when the user wants the whole Agent integration turned off or wants every managed skill removed from it. If they only want one skill or preset removed while keeping the Agent available for future deployments, use `skills undeploy` or `presets undeploy` instead.

## Typical workflows

### "Find me a skill for X" / "Install a skill that does X"

1. `skills search "X" --limit 5` — show the top 1–3 hits with install counts and source.
2. If a clear winner: `skills install <install_ref>`.
3. If ambiguous: ask the user to pick.
4. Deploy it to the agent(s) the user requested with `skills deploy`.
5. `skills status <name>` to confirm the library and deployment state.

### "What skills do I have?"

```bash
"$SM" --json skills list
```

The `preset_ids`, `presets`, `deployed_to`, `tags`, and `source_type` fields are usually the most informative. The legacy `enabled` field is not deployment state.

### "Pull in the skills already installed in my agent directories"

1. `skills adopt ~/.claude/skills --dry-run` (and any other agent dirs the user mentions) — show the candidate list.
2. After user confirms: `skills adopt ~/.claude/skills`.
3. For any adopted skill where the user knows the original repo, restore the update link with `skills set-source <skill> --git-url ... --subpath ...`.

### "Update everything"

```bash
"$SM" skills check --all     # see what has upstream changes
"$SM" skills update --all    # apply
```

Report which skills actually refreshed (`refreshed: true` in the JSON) vs which were already up-to-date.

## Pitfalls

- **Install succeeded but skill doesn't appear in the agent** → install defaults to library-only. Use `skills deploy <skill> --agent <key>`.
- **Preset membership changed but agent files did not** → membership is organization only. Follow with `presets deploy` or `skills deploy` when the user also asked to make it visible.
- **No active preset** only affects legacy `skills sync` / `presets apply`; additive deploy commands do not require one.
- **Adopted skills can't be `update`d from git** → `npx skills add` and manual `git clone` don't leave source metadata, so adopt has to treat them as `local`. Re-point them with `skills set-source`. Do **not** reach for `adopt --git-url` here: adopt only ever creates new library entries, and it fails *late* — `--dry-run` returns `ok: true` with the skill sitting in `skipped`, and only the real run errors with `--git-url requires exactly one adoptable skill, found 0`. Do **not** remove-then-reinstall either — that drops the skill id, and with it the tags, preset membership and every per-agent deployment.
- Use `--dry-run` before bulk remove, tag delete, preset delete, deploy, or undeploy operations. Use `check` before `update`.