Skill Workshop

Skill Workshop is OpenClaw's governed path for creating and updating workspace skills. Instead of agents writing directly to SKILL.md files, they propose skills through a review queue that requires human approval before activation.

Introduced in v2026.6.1, Skill Workshop adds audit trails, security scanning, hash binding, and rollback to the skill creation process.

info

Skill Workshop manages workspace skills only. It does not modify bundled skills, plugin skills, ClawHub-installed skills, or system skills.

Why Skill Workshop?

	Before v2026.6.1	With Skill Workshop
Creation	Agents write directly to `SKILL.md`	Agents propose via `skill_workshop` tool
Review	None — skills go live immediately	Mandatory review queue
Audit trail	Filesystem history only	Explicit versioning + metadata
Safety	Bad skills become permanent	Review gate blocks bad patterns
Support files	Manual management	Scanned, hashed, versioned with proposal
Rollback	Manual intervention	Metadata-backed recovery
Governance	None	Tracked approvals, rejections, quarantines

Quick Start

# 1. Create a proposal
openclaw skills workshop propose-create \
  --name "daily-standup" \
  --description "Run a daily standup: what I did, what I'll do, blockers" \
  --proposal ./PROPOSAL.md

# 2. Review it
openclaw skills workshop inspect daily-standup-001

# 3. Apply it (activates the skill)
openclaw skills workshop apply daily-standup-001

That's the core loop: propose → review → apply.

Proposal Lifecycle

Every proposal moves through a defined set of states:

             ┌─→ applied (live skill)
             │
pending ─────┼─→ rejected
             │
             ├─→ quarantined (held for deeper review)
             │
             └─→ stale (live skill changed after proposal was created)

State	Description	Terminal?
Pending	Created or revised, ready for review	No — can revise, apply, reject, or quarantine
Applied	Approved and written to live skill files	Yes
Rejected	Declined by operator, with optional reason	Yes
Quarantined	Flagged for deeper review (security concerns, etc.)	Yes
Stale	Live skill was modified after the proposal was created — hash no longer matches	Yes

Only pending proposals can be revised. Terminal states require creating a new proposal.

CLI Commands

Create a Proposal

openclaw skills workshop propose-create \
  --name "audit-log-parser" \
  --description "Parse and analyze audit logs for security events" \
  --proposal ./PROPOSAL.md

Creates a new skill proposal. Fails immediately if a skill with that name already exists (no-clobber protection).

Update an Existing Skill

openclaw skills workshop propose-update <skill-name> \
  --proposal ./PROPOSAL.md

Creates a proposal to update a live skill. The proposal binds to the skill's current hash — if the skill changes before you apply, the proposal goes stale.

Add Support Files

Proposals can include templates, scripts, examples, and references:

openclaw skills workshop propose-create \
  --name "config-validator" \
  --proposal ./my-proposal/PROPOSAL.md \
  --proposal-dir ./my-proposal

Expected directory structure:

my-proposal/
├── PROPOSAL.md          # Required: the skill content
├── assets/              # Media files, icons
├── examples/            # Usage examples, sample outputs
├── references/          # Documentation, links
├── scripts/             # Helper scripts
└── templates/           # Reusable templates

Constraints: max 64 files, 256 KB each, 2 MB total. No executables, no absolute paths, no path traversal, no null bytes.

Revise a Pending Proposal

openclaw skills workshop revise <proposal-id> \
  --proposal ./PROPOSAL-v2.md

Revises in place — auto-increments the version number and preserves history.

List Proposals

openclaw skills workshop list
openclaw skills workshop list_quarantine

Inspect a Proposal

openclaw skills workshop inspect <proposal-id>

Shows the full proposal content, revision history, scanner verdicts, and metadata.

Review Actions

# Approve and activate
openclaw skills workshop apply <proposal-id>

# Decline
openclaw skills workshop reject <proposal-id> \
  --reason "Duplicate functionality — use the existing deploy skill"

# Flag for deeper review
openclaw skills workshop quarantine <proposal-id> \
  --reason "Suspicious external API calls; needs manual security review"

Scanner Gating

When you apply a proposal, OpenClaw reruns security scanning before writing the live skill. Three independent scanners check the skill:

Scanner	What It Does
OpenClaw static analysis	Built-in pattern matching for known bad patterns
VirusTotal	Threat detection against the VirusTotal database
NVIDIA SkillSpector	AI-assisted semantic analysis for hidden instructions and risky code paths

ClawScan Verdicts

The combined scan produces one of three verdicts:

Verdict	Meaning	Action
Clean	Passed all checks	Safe to apply
Suspicious	Unclear disclosure, overbroad authority, risky defaults, or high blast radius	Review carefully before applying
Malicious	Confirmed threat	Automatically quarantined

If the scanner finds critical issues during apply, the proposal is automatically moved to quarantined instead of being activated.

Hash Binding

Update proposals bind to the target skill's current content hash. This prevents conflicts:

If the live skill changes while a proposal is pending, the proposal becomes stale
Stale proposals cannot be revised or applied
The agent must create a fresh proposal against the new skill state

This guarantees you never accidentally overwrite changes made between proposal creation and review.

Rollback

When a proposal is applied, rollback metadata is written before the live skill files are modified. If something goes wrong after apply:

The rollback record contains the previous skill content and hash
Recovery can restore the pre-apply state
All version history is preserved for audit

The `skill_workshop` Agent Tool

Agents interact with Skill Workshop through a dedicated tool — they cannot modify proposals via direct filesystem operations.

{
  "tool": "skill_workshop",
  "action": "create",       // create, update, revise, list, inspect
  "name": "skill-name",
  "description": "Concise description (max 160 bytes)",
  "content": "Full skill content",
  "supportFiles": { }       // Optional
}

By default, agent-initiated apply, reject, and quarantine actions show an approval prompt. Configure this with approvalPolicy.

Control UI

The dashboard provides two views for reviewing proposals:

Board View

The full workshop interface:

Searchable proposal list filtered by state (pending, applied, rejected, stale, quarantined)
Inspectable support file previews
Revision diffs showing what changed across versions
Scanner verdict display

Today View

A fast-pass for one-at-a-time decisions:

Shows the next pending proposal
Presents a focused question: "Should this become part of your skill set?"
Accept, reject, or quarantine with one click
Moves to the next proposal automatically

Both views support session routing — start a review, hand it off, pick it up later.

Configuration

~/.openclaw/openclaw.json
{
  "skills": {
    "workshop": {
      "approvalPolicy": "pending",    // "pending" (prompt) or "auto" (skip prompt)
      "maxPending": 50,               // Max pending + quarantined proposals
      "maxSkillBytes": 40000,         // Proposal body size limit (bytes)
      "autonomous": {
        "enabled": false              // Auto-create proposals from conversations
      }
    }
  }
}

Setting	Default	Description
`approvalPolicy`	`"pending"`	`"pending"` requires operator approval for agent-initiated apply/reject/quarantine. `"auto"` skips the prompt.
`maxPending`	`50`	Maximum number of pending + quarantined proposals per workspace. Prevents queue overflow.
`maxSkillBytes`	`40000`	Maximum size of the proposal body in bytes.
`autonomous.enabled`	`false`	When `true`, agents can auto-create proposals from successful conversation turns. Experimental — leave off for most setups.

warning

Set approvalPolicy: "pending" in production. The "auto" setting should only be used in CI environments, isolated sandboxes, or fully trusted deployments.

Storage

Proposals live in ~/.openclaw/skill-workshop/:

skill-workshop/
├── proposals.json                 # Fast listing index
└── proposals/
    └── <proposal-id>/
        ├── proposal.json          # Canonical proposal record
        ├── PROPOSAL.md            # Pending skill content
        ├── rollback.json          # Recovery metadata (written on apply)
        ├── assets/
        ├── examples/
        ├── references/
        ├── scripts/
        └── templates/

Practical Examples

Agent-Driven Skill Creation

Ask your bot to create a skill through chat:

Create a skill called "pr-review" that reviews GitHub pull requests.
It should check for: code style, security issues, test coverage,
and documentation. Output a structured review with severity levels.

The agent calls skill_workshop with action: create → you see it in the proposal queue → review and apply from the dashboard or CLI.

# Agent creates initial proposal
# You inspect it
openclaw skills workshop inspect pr-review-001

# Not quite right — ask the agent to revise
# "Revise the pr-review skill to also check for breaking API changes"

# Agent calls skill_workshop with action: revise
# Version auto-increments, full history preserved

# Happy with v3 — apply it
openclaw skills workshop apply pr-review-001

Bulk Review with Today View

If your agents have been busy creating proposals:

Open the dashboard → Skill Workshop → Today View
Review each proposal one by one
Apply the good ones, reject duplicates, quarantine anything suspicious
Each decision takes seconds

Security Audit of Existing Skills

# Verify a skill against ClawHub's security database
openclaw skills verify deploy-helper --card

# Review all quarantined proposals
openclaw skills workshop list_quarantine

Best Practices

For production setups:

Keep approvalPolicy: "pending" — always review before activating
Leave autonomous.enabled: false unless you specifically need auto-proposals
Review scanner verdicts before applying — don't ignore "Suspicious" findings
Check support files for sensitive data or executables
Keep skill descriptions clear and under 160 bytes
Don't delete quarantined proposals without logging why

For skill authors:

Write clear, single-purpose skills — one skill, one job
Include examples in the examples/ directory
Test your skill as a proposal before applying to production
Use the revision workflow instead of creating new proposals for the same skill

Skill Workshop

Why Skill Workshop?

Quick Start

Proposal Lifecycle

CLI Commands

Create a Proposal

Update an Existing Skill

Add Support Files

Revise a Pending Proposal

List Proposals

Inspect a Proposal

Review Actions

Scanner Gating

ClawScan Verdicts

Hash Binding

Rollback

The `skill_workshop` Agent Tool

Control UI

Board View

Today View

Configuration

Storage

Practical Examples

Agent-Driven Skill Creation

Iterative Refinement

Bulk Review with Today View

Security Audit of Existing Skills

Best Practices

See Also

Why Skill Workshop?​

Quick Start​

Proposal Lifecycle​

CLI Commands​

Create a Proposal​

Update an Existing Skill​

Add Support Files​

Revise a Pending Proposal​

List Proposals​

Inspect a Proposal​

Review Actions​

Scanner Gating​

ClawScan Verdicts​

Hash Binding​

Rollback​

The skill_workshop Agent Tool​

Control UI​

Board View​

Today View​

Configuration​

Storage​

Practical Examples​

Agent-Driven Skill Creation​

Iterative Refinement​

Bulk Review with Today View​

Security Audit of Existing Skills​

Best Practices​

See Also​

Why Skill Workshop?

Quick Start

Proposal Lifecycle

CLI Commands

Create a Proposal

Update an Existing Skill

Add Support Files

Revise a Pending Proposal

List Proposals

Inspect a Proposal

Review Actions

Scanner Gating

ClawScan Verdicts

Hash Binding

Rollback

The `skill_workshop` Agent Tool

Control UI

Board View

Today View

Configuration

Storage

Practical Examples

Agent-Driven Skill Creation

Iterative Refinement

Bulk Review with Today View

Security Audit of Existing Skills

Best Practices

See Also