crewship skill

Skills are reusable instruction playbooks that get injected into an agent’s system prompt as a [SKILLS AVAILABLE] block and materialised on disk in each CLI’s discovery convention (.claude/skills/, .opencode/skills/, .factory/skills/, .agents/skills/, .cursor/rules/). The CLI is the canonical authoring + iteration path; the web UI is a surface over the same API.

crewship skill <subcommand> [flags]

Subcommands at a glance

Command	What it does
`init`	Generate a SKILL.md scaffold on disk (offline; no API call).
`list`	List skills in the workspace, with optional filters.
`get`	Show skill metadata.
`export`	Recover a previously-imported SKILL.md from the workspace registry.
`import`	Upload a SKILL.md (single URL/file) or walk a whole git repo.
`create`	LLM-author a SKILL.md via Anthropic Messages (requires API key).
`delete`	Remove a skill from the workspace registry.
`assign`	Attach a skill to one agent, several agents, or a whole crew.
`unassign`	Detach the same way.
`proposed list`	List skills auto-promoted from memory awaiting operator approval.
`proposed approve`	Import a staged proposal via the canonical importer.
`proposed reject`	Discard a staged proposal (deletes the `.proposed/` file).

The CLI authoring loop

The intended flow for a custom skill is:

# 1. Scaffold a starter SKILL.md
crewship skill init pdf-cleanup --category DEVOPS \
  --description "Use when the user asks to strip metadata from a PDF."

# 2. Edit the scaffold — fill in Steps, Output format, Guardrails
$EDITOR ./pdf-cleanup/SKILL.md

# 3. Upload it
crewship skill import --file ./pdf-cleanup/SKILL.md

# 4. Assign to one or more agents
crewship skill assign pdf-cleanup --to-crew engineering

# 5. Iterate: pull the current version back, edit, re-upload
crewship skill export pdf-cleanup --output ./
$EDITOR ./pdf-cleanup.md
crewship skill import --file ./pdf-cleanup.md

init is intentionally offline — it doesn’t talk to the server, so it works for users who only have a Claude Code OAuth token and cannot reach the Anthropic Messages API directly (the path crewship skill create needs).

`crewship skill init`

Write a SKILL.md scaffold to disk so you can edit it locally and upload.

crewship skill init <slug> [flags]

Flag	Default	Notes
`--category`	`CUSTOM`	One of `CODING DATA DEVOPS WRITING RESEARCH PM DESIGN SUPPORT SECURITY FINANCE OPS AUTOMATION SALES CUSTOM`.
`--description`	(empty)	One-line trigger; should start with `Use when...`, `Useful for...`, or `To <verb>...`. When omitted, the scaffold writes a `TODO:` placeholder into the frontmatter `description:` field — fill it in before `crewship skill import`, which validates the SKILL.md server-side.
`--license`	`MIT`	SPDX identifier — must be on the allowlist below for `skill import` to accept it later.
`--output`	`./<slug>/`	Output directory for `<dir>/SKILL.md`.
`--force`	`false`	Overwrite an existing SKILL.md at the destination.

The scaffold writes the canonical body sections (When to use, Steps, Output format, Guardrails, optional Verification) and frontmatter fields the parser accepts.

`crewship skill list`

crewship skill list [flags]

Flag	Notes
`--category`	Filter by category.
`--source`	One of `BUNDLED CUSTOM GENERATED MARKETPLACE MANAGED`.
`--vendor`	Vendor namespace (e.g. `anthropic`, `community`).
`--maturity`	One of `OFFICIAL CURATED COMMUNITY EXPERIMENTAL`.
`--runtime`	One of `INSTRUCTIONS SCRIPT MCP HYBRID`.
`--search`	Substring match on name / display_name / description.
`--installed`	Only skills with at least one `agent_skills` row.
`--installed-for=<agent>`	Only skills assigned to a specific agent.

Examples:

crewship skill list --maturity OFFICIAL
crewship skill list --vendor anthropic --category DESIGN
crewship skill list --installed-for viktor
crewship skill list --search pdf

`crewship skill get`

Show metadata for one skill — name, slug, category, version, source, author, tool count, created-at. The markdown description body renders below the metadata block via glamour for human formats; JSON / YAML / quiet outputs include it raw inside the serialized struct.

crewship skill get <slug-or-id>
crewship skill get pdf-cleanup --format json | jq '.tool_count'

Backed by GET /api/v1/skills/{id} after resolving <slug-or-id> to the skill’s id.

`crewship skill export`

Recover the SKILL.md body from the workspace and write it to disk or stdout.

crewship skill export <slug-or-id> [--output PATH]

Without --output, prints to stdout (pipe-friendly). With --output set to a directory, writes <slug>.md inside it.

# Print to stdout
crewship skill export pdf-cleanup

# Write to a directory
crewship skill export pdf-cleanup --output ./

`crewship skill import`

Import a single SKILL.md (URL or file) or walk a whole git repo.

# Single SKILL.md from a URL or local file
crewship skill import https://raw.githubusercontent.com/owner/repo/main/skills/my-skill/SKILL.md
crewship skill import --file ./SKILL.md

# Whole git repo (--depth 1, walks for **/SKILL.md, license-gated)
crewship skill import --repo https://github.com/anthropics/skills
crewship skill import --repo https://github.com/owner/skills --paths 'skills/*' --dry-run
crewship skill import --repo https://github.com/owner/skills --unsafe-license

Flag	Notes
`--file`	Path to a local SKILL.md (single skill).
`--repo`	Git URL — switches to bulk mode.
`--ref`	Git ref (branch/tag); only with `--repo`.
`--paths`	Glob filters relative to repo root, e.g. `skills/*`.
`--vendor`	Override vendor namespace (defaults to `community`).
`--unsafe-license`	Skip the SPDX license allowlist (use with caution).
`--dry-run`	Walk + parse, don’t write.

What runs on every import

SSRF check — URL/host must be public; private IPs blocked (single + bulk).
License gate — frontmatter license: is matched against the SPDX allowlist below.
Prompt-injection scanner — body scanned for “ignore previous instructions”, role hijack, large base64 blobs, etc. Status stored in scan_status (CLEAN / FLAGGED).
Size cap — single fetch ≤ 512 KB; bulk per-file same; bulk total ≤ 500 SKILL.md files.
BUNDLED protection — re-importing a slug already owned by a BUNDLED skill is refused.

SPDX allowlist

MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, ISC, CC0-1.0, MPL-2.0, Unlicense, 0BSD. Pass --unsafe-license to bypass for one batch (audited via the license column).

`crewship skill create`

LLM-authored SKILL.md via Anthropic Messages.

crewship skill create --slug pdf-cleanup \
  --prompt "Help users sanitise PDFs: strip metadata, remove embedded JS, flatten forms"

Requires an active Anthropic credential of type API_KEY in the workspace (Settings → Credentials). An OAuth token from claude /login is not sufficient (it’s a Bearer for claude.ai, not the Messages API). Users without an API key should reach for skill init instead.

Flag	Default	Notes
`--slug`	required	Kebab-case identifier.
`--prompt`	required	Free-form description of the skill’s behaviour.
`--model`	`claude-sonnet-4-6`	LLM model override.
`--print`	`false`	Print generated SKILL.md to stdout instead of summary.

`crewship skill delete`

Remove a skill from the workspace registry. Cascades to agent_skills.

crewship skill delete <slug-or-id>          # interactive confirm
crewship skill delete <slug-or-id> --yes    # skip confirm (--force also accepted)
crewship skill rm <slug-or-id>              # alias

Cascades to agent_skills — every agent loses the skill. BUNDLED skills cannot be deleted (they re-seed on every server start).

`crewship skill assign`

# One agent (positional)
crewship skill assign my-skill viktor

# Multiple agents
crewship skill assign my-skill --to-agents viktor,nela,martin

# Whole crew
crewship skill assign my-skill --to-crew engineering

The fan-out variant is idempotent: agents that already have the skill are treated as success, and per-agent failures are reported but don’t abort the rest.

`crewship skill unassign`

Same target shapes as assign:

crewship skill unassign my-skill viktor
crewship skill unassign my-skill --to-agents viktor,nela
crewship skill unassign my-skill --to-crew engineering

After unassign, the next time the agent runs, the orchestrator prunes the skill folder from the agent’s container (.claude/skills/<slug>/, .cursor/rules/<slug>.mdc, etc.) so a CLI auto-discovery walker can no longer pick it up.

`crewship skill proposed`

Operator-side HITL surface for the memory→Skills bridge. When the consolidator sees a learned rule with sustained recall (default ≥ 10 recall events, composite score ≥ 0.85) it stages an Anthropic-format SKILL.md under .memory/{crew-slug}/topics/.proposed/skill-<slug>.md. These three verbs let an operator inspect and decide.

Each proposed verb requires OWNER, ADMIN, or MANAGER role.

`crewship skill proposed list`

crewship skill proposed list --crew <crew-id>
crewship skill proposed list --crew <crew-id> --format json

Flag	Required	Description
`--crew`	yes	Crew id whose `.proposed/` dir to scan. The CLI URL-encodes this value so any future id-scheme change with special chars stays safe.
`--format`	no	`text` (default) or `json`.

Output columns: FILE, NAME, CATEGORY, DESCRIPTION (truncated to 60 chars), QUALITY (ok when the server didn’t flag it). Backed by GET /api/v1/skills/proposed?crew_id=<id>.

`crewship skill proposed approve`

Run the staged file through the canonical skill importer — same path as URL-based imports, so the SPDX license check and prompt-injection scan apply. On success the staging file is deleted and an EntryMemorySkillApproved journal entry fires.

crewship skill proposed approve --crew <crew-id> --file skill-deploy-friday.md

Flag	Required	Description
`--crew`	yes	Crew id.
`--file`	yes	Staged file name as listed (e.g. `skill-deploy-friday.md`).

Backed by POST /api/v1/skills/proposed/approve. Output: approved <file> -> skill <id> (<slug>, created|updated).

`crewship skill proposed reject`

Discard the staging file without importing. Idempotent — rejecting an already-deleted file returns success, and an EntryMemorySkillRejected journal entry fires either way.

crewship skill proposed reject --crew <crew-id> --file skill-noise.md

Flag	Required	Description
`--crew`	yes	Crew id.
`--file`	yes	Staged file name.

Backed by POST /api/v1/skills/proposed/reject.

Per-CLI filesystem materialisation

When a skill is assigned, the orchestrator writes its SKILL.md into the discovery convention each adapter knows. This way the same skill works for whatever CLI the agent is configured with:

CLI	Path inside container
Claude Code	`.claude/skills/<slug>/SKILL.md`
OpenAI Codex CLI	`.agents/skills/<slug>/SKILL.md`
OpenCode	`.opencode/skills/<slug>/SKILL.md`
Factory Droid	`.factory/skills/<slug>/SKILL.md`
Cursor	`.cursor/rules/<slug>.mdc` (flat .mdc)
Gemini CLI / other adapters	`[SKILLS AVAILABLE]` block only (no FS)

The system-prompt block is the authoritative surface — filesystem layout is the bonus discovery path for adapters that walk it natively.

​crewship skill

​Subcommands at a glance

​The CLI authoring loop

​crewship skill init

​crewship skill list

​crewship skill get

​crewship skill export

​crewship skill import

​What runs on every import

​SPDX allowlist

​crewship skill create

​crewship skill delete

​crewship skill assign

​crewship skill unassign

​crewship skill proposed

​crewship skill proposed list

​crewship skill proposed approve

​crewship skill proposed reject

​Per-CLI filesystem materialisation

crewship skill

Subcommands at a glance

The CLI authoring loop

`crewship skill init`

`crewship skill list`

`crewship skill get`

`crewship skill export`

`crewship skill import`

What runs on every import

SPDX allowlist

`crewship skill create`

`crewship skill delete`

`crewship skill assign`

`crewship skill unassign`

`crewship skill proposed`

`crewship skill proposed list`

`crewship skill proposed approve`

`crewship skill proposed reject`

Per-CLI filesystem materialisation