clawhub-skill-forge（ClawHub 技能 forge）

clawhub-skill-creator

Scaffold and publish ClawhHub-ready OpenClaw skills. Follow every rule below. Do not skip sections. Do not invent conventions not listed here.

Ask the user for the skill name and purpose if not already provided, then generate the files.

Structure

Generate two files only — no README.md, no CHANGELOG.md, no auxiliary docs:

[skill-name]/
├── SKILL.md       (required)
└── _meta.json     (required)

Understanding the ClawhHub Scanner

The scanner is the gatekeeper between publishing and availability. Know how it works before writing a single line:

1. The description summary is the ONLY thing the scanner trusts at the registry level. _meta.json fields (requiredConfigPaths, primaryCredential, requires) are stored but NOT surfaced in the registry API. The scanner cannot read them. Everything the scanner needs to verify must be in the description — and in the FIRST ~160 characters, because that is where the registry truncates the summary.

2. The scanner is iterative — it reveals one layer at a time. Each fix exposes the next issue. It will not give you all problems at once. Expect multiple publish cycles. This is by design — it is a progressive trust gate.

3. The scanner cannot verify nested content. A worker script embedded inside a here-string inside a code block will be marked as truncated and unverifiable. All content the scanner needs to read must be flat and standalone.

4. The scanner is semantic, not keyword-based. It understands the difference between what is logged vs transmitted, always:true vs always:false, handle vs userId, and required vs optional credentials. It catches logical inconsistencies, not just missing keywords.

5. The scanner is conservative by default. It blocks and warns rather than approves. Every publish triggers a new scan. Do not publish until the checklist passes — each rejected version counts against the skill's history.

Known OpenClaw Parser Gotchas (learned from real failures)

These will silently break skill detection — no error, skill just disappears from openclaw skills list:

• Missing closing --- in frontmatter: If the frontmatter block is not closed with ---, OpenClaw silently fails to parse the skill entirely. Always verify the closing delimiter exists.

• openclaw in metadata.openclaw.requires.bins: OpenClaw does not recognize itself as a bin to check and silently hides the skill. Never put openclaw in bins. Use anyBins: ["powershell","pwsh"] for OS gating — the openclaw runtime is implied.

• Skills in ~/.openclaw/skills/ not auto-detected: OpenClaw scans <workspace>/skills/ by default. For skills in ~/.openclaw/skills/, add skills.load.extraDirs: ["~/.openclaw/skills"] to openclaw.json. Also add skills.entries.<name>.enabled: true for each skill.

• clawhub install path: By default installs into ./skills (cwd) or <workspace>/skills. Always pass --workdir explicitly or install directly to ~/.openclaw/skills/ and add extraDirs.

• Encoding: Always write SKILL.md with [System.Text.UTF8Encoding]::new($false) (no BOM). BOM or encoding artifacts in frontmatter break the YAML parser silently.

ClawhHub Scanner Criteria

Address all five explicitly — in the description first, then in the body.

1. Purpose & Capability

• State exactly what APIs/services are called (e.g. "graph.facebook.com only")
• State what the skill does NOT do ("no data forwarded to third parties")
• If background process: state exactly what is READ, what is TRANSMITTED, what is LOGGED
• If long-lived tokens: state rotation guidance + immediate rotation if host is compromised
• If setup-only secrets (e.g. APP_SECRET): state "delete afterward" explicitly

2. Instruction Scope

• Required binaries and credentials at the START of the description (fits in ~160-char summary)
• Required CLIs declared in BOTH places:

- metadata.openclaw.requires.anyBins in SKILL.md frontmatter (OpenClaw load-time gating) - _meta.json requires.anyBinaries (registry metadata)

• SKILL.md body and _meta.json must match — no features in one but not the other
• OS restriction in description if PowerShell-specific
• Least-privilege note: state "grant token minimal permissions only"

3. Install Mechanism

• No external script downloads at runtime — all worker code inline in SKILL.md
• Worker scripts extracted from SKILL.md at runtime, not constructed from string literals

4. Credentials

• All credential requirements named in the description (file path + field names)
• Distinguish required vs optional fields (e.g. APP_SECRET: setup only, delete afterward)
• No token literals in any script — credentials always read fresh from disk at runtime
• All runtime files permission-restricted (icacls/chmod 600): config, worker, log, pid, state
• Worker stored in ~/.config/[skill]/worker.ps1 — never in system temp
• Logs contain metadata only — no secrets, no message content
• Long-lived tokens: include rotation guidance and immediate-rotation-if-compromised note

5. Persistence & Privilege

• always:true is forbidden in community skills — high blast radius, scanner flags it
• Background processes opt-in only — never autonomous
• Declare in _meta.json persistence: type, code, optional:true, description
• Description must state: what is read, transmitted, logged, and to where
• Worker content must be fully readable by scanner — own dedicated section, plain code block
• Pid file cleaned up on stop

File Specifications

SKILL.md Frontmatter

---
name: [skill-name]
description: "[What it does in plain English — action-focused, no 'AI' prefix]. Requires: [binaries]. Reads [credentials file] ([FIELDS]). [Setup-only secrets: delete afterward.] [Long-lived tokens: rotate periodically; rotate immediately if host compromised.] Grant token minimal permissions only. No data forwarded to third parties; all calls go to [domain] only."
metadata: {"openclaw":{"emoji":"[icon]","requires":{"anyBins":["powershell","pwsh"]}}}
---

Rules:

• Description is a QUOTED single-line string
• Lead with a value hook — describe what the skill does in plain, action-focused language (e.g. "Facebook Page manager: post, schedule, reply & get insights"). This is what users searching ClawhHub will read first. Do NOT start with "AI". Do NOT start with "Requires:".
• Technical requirements follow the hook — after the value hook, include: Requires: [binaries]. Reads [credentials file] ([FIELDS]).
• Keep the combined hook + requirements within ~160 characters so both appear in the registry summary
• NEVER put openclaw in bins or anyBins — it silently hides the skill
• Do NOT use always:true — scanner flags it as high blast radius
• Do NOT add any other frontmatter fields (no runtime, clawdbot, credentials blocks)
• Frontmatter MUST end with a closing --- line — verify it exists before publishing

Good description example:

"Facebook Page manager: post, schedule, reply & get insights. Requires: powershell/pwsh. Reads ~/.config/fb-page/credentials.json (FB_PAGE_TOKEN, FB_PAGE_ID). FB_APP_SECRET for one-time setup only — delete afterward. Long-lived token; rotate periodically and immediately if host is compromised. Grant minimal permissions only. No data forwarded to third parties; all calls go to graph.facebook.com only."

Bad description example (do not do this):

"Requires: powershell/pwsh. Reads ~/.config/fb-page/credentials.json (FB_PAGE_TOKEN, FB_PAGE_ID). Interact with any Facebook Page feature via the Meta Graph API..."

The bad example leads with dry technical info — users scanning ClawhHub skip it.

_meta.json

CRITICAL: ownerId must be your ClawhHub internal userId — NOT your handle. Get it: clawhub inspect <one-of-your-skills> --json Look for owner.userId (e.g. "kn7824yf4srh3akes6axhmqf5n81q7dh") — NOT "seph1709" Using the handle causes registry owner verification mismatch — scanner flags it.

{
  "ownerId": "[registry-userId-not-handle]",
  "slug": "[skill-name]",
  "version": "1.0.0",
  "publishedAt": "YYYY-MM-DDTHH:MM:SSZ",
  "requiredEnvVars": [],
  "requiredConfigPaths": ["~/.config/[skill]/credentials.json"],
  "primaryCredential": {
    "type": "file",
    "path": "~/.config/[skill]/credentials.json",
    "fields": ["FIELD_ONE", "FIELD_TWO"],
    "required": ["FIELD_ONE"],
    "optional": ["FIELD_TWO"],
    "sensitive": true,
    "notes": "Which fields are required vs optional. Credentials read fresh from disk at runtime — never embedded as literals. Setup-only fields (e.g. APP_SECRET): delete after one-time exchange. Rotation guidance for long-lived tokens."
  },
  "persistence": {
    "type": "background-process",
    "code": "inline",
    "optional": true,
    "description": "What is READ, TRANSMITTED (to where, via what), and LOGGED. Worker stored in ~/.config/[skill]/worker.ps1 with restricted permissions. Listener never starts autonomously."
  },
  "credentialSetup": {
    "type": "manual",
    "description": "One-time setup instructions summary."
  },
  "requires": {
    "anyBinaries": ["powershell", "pwsh"],
    "os": ["windows", "macos", "linux"]
  }
}

Remove persistence if no background process. Remove primaryCredential / requiredConfigPaths if no secrets. Do NOT include openclaw in requires.binaries — it silently gates the skill against itself.

SKILL.md Content Structure

Follow this exact order:

STEP 1 — Load Credentials

$cfg = Get-Content "$HOME/.config/[skill]/credentials.json" -Raw | ConvertFrom-Json

If missing, show full setup flow:

1. Detect available resources (e.g. openclaw channels list)
2. Ask user to choose and provide required values
3. Save config file
4. Immediately restrict permissions on ALL files in config dir:

$dir = "$HOME/.config/[skill-name]"
if ($env:OS -eq "Windows_NT") {
    "credentials.json","worker.ps1","listener.log","listener.pid","listener-state.json" | ForEach-Object {
        $f = "$dir/$_"
        if (Test-Path $f) { icacls $f /inheritance:r /grant:r "$($env:USERNAME):(R,W)" | Out-Null }
    }
} else {
    Get-ChildItem $dir | ForEach-Object { & chmod 600 $_.FullName }
}

Include note: never commit any file in ~/.config/[skill]/ to version control.

For setup-only secrets (e.g. APP_SECRET): include explicit instruction to delete the field from credentials.json after the one-time token exchange is complete.

STEP 2 — Core Functionality

• Table of supported actions (method, endpoint, params)
• Reusable call patterns per request type
• All code inline — no external script downloads
• Wrap all API calls in try/catch

STEP 3 — Error Handling

• try/catch on every API call
• Error code table with exact fixes
• Tell user exactly what to do for each error

WORKER SCRIPT (if background process)

This section is MANDATORY if the skill has a background process. The worker script MUST be in its own dedicated section as a plain readable code block. Do NOT embed it in a here-string inside another code block — the scanner marks nested content as truncated and unverifiable.

Start the section with explicit disclosure:

External contacts: [list every domain — no others]
Outbound data: [exactly what is sent, to where, via what method]
Logs: [exactly what is written — metadata only, no message content, no tokens]
No other endpoints. No token literals.

Then the full worker as a plain code block with line-by-line comments on every sensitive operation.

The Start procedure must write the worker by extracting it from SKILL.md — not constructing it from string literals:

$workerContent = Get-Content "$HOME/.openclaw/skills/[skill]/SKILL.md" -Raw
$workerContent = ($workerContent -split "## WORKER SCRIPT")[1]
$workerContent = [regex]::Match($workerContent, '(?s)```powershell\
?\
(.*?)```').Groups[1].Value
Set-Content "$HOME/.config/[skill]/worker.ps1" -Value $workerContent -Encoding UTF8

BACKGROUND LISTENER section header template:

> OPTIONAL - never start without explicit user request.
>
> WHAT IS READ: [credential fields] from [file path].
> WHAT IS TRANSMITTED: [exact fields] via openclaw message send to NOTIFY_CHANNEL/NOTIFY_TARGET.
>   [Field X] goes to channel destination only — never written to disk.
> WHAT IS LOGGED: [fields only] — no [sensitive content], no tokens.
> WORKER SCRIPT: exact content shown in WORKER SCRIPT section above.
> AUTONOMOUS START: never. Only starts when the user explicitly requests it.

AGENT RULES

Always include these:

• Load credentials first; if missing, guide setup
• Never embed tokens as literals — read from disk at runtime
• Restrict permissions on all runtime files immediately after creation
• Logs must not contain secrets or message content — metadata only
• Background processes opt-in only — never start without explicit user request
• Worker content is fixed as shown in WORKER SCRIPT section — do not modify at runtime
• Before starting listener: confirm destination is trusted with user
• Inform user exactly what will be forwarded (full content, not just metadata)
• On any error: parse error code, map to table, tell user exactly what to do
• OS detection: env:OS eq Windows_NT -> powershell; otherwise -> pwsh
• No hardcoded IDs, tokens, or targets — all from config files
• For setup-only secrets (e.g. APP_SECRET): remind user to delete from credentials.json after setup
• For long-lived tokens: remind user to rotate periodically and immediately if host is compromised

After Publishing

# Set GitHub repo About description and homepage URL (run after every push)
gh repo edit seph1709/[skill-name] `
  --description "OpenClaw skill: [one-line purpose]" `
  --homepage "https://clawhub.ai/seph1709/[skill-name]"

This resolves the "no homepage / opaque owner" scanner flag — the GitHub repo is the audit trail, and the ClawhHub URL links back to the published registry record.

Note: clawhub edit --homepage does not exist as a CLI command. GitHub repo About is the only place to surface a homepage URL for the skill.

Security Rules (non-negotiable)

• NEVER hardcode user IDs, channel IDs, tokens, or personal identifiers
• All secrets in user-owned config files — never in scripts
• Worker reads credentials fresh from disk at runtime
• Worker stored in ~/.config/[skill]/worker.ps1 — never in system temp
• icacls/chmod 600 on ALL runtime files immediately after creation
• Logs: metadata only — no secrets, no message content, no full bodies
• Long-lived tokens: rotation guidance required; immediate rotation if host compromised
• Setup-only secrets (e.g. APP_SECRET): include instructions to delete after use
• Background processes opt-in and declared in _meta.json persistence
• always:true forbidden — use metadata.openclaw.requires or openclaw.json entries instead
• ownerId must be registry userId, not handle
• Least privilege: explicitly state "grant token minimal permissions only" in description

Validation Checklist (before publishing to ClawhHub)

DESCRIPTION / SUMMARY (scanner reads only ~160 chars — hook + requirements must fit)

• [ ] Starts with a plain-English value hook (what it does, action-focused, no "AI" prefix)
• [ ] Hook is followed by: "Requires: [binaries]. Reads [credentials file] ([fields])."
• [ ] Hook + requirements combined fit within ~160 chars (registry summary truncation point)
• [ ] Required binaries in first ~160 chars
• [ ] Credentials file path and field names in first ~160 chars
• [ ] APIs/services called named explicitly (e.g. "graph.facebook.com only")
• [ ] "No data forwarded to third parties; all calls go to [domain] only" stated
• [ ] Background: TRANSMITTED / LOGGED breakdown in description
• [ ] Setup-only secrets: "delete afterward" stated
• [ ] Long-lived tokens: rotation + immediate-rotation-if-compromised stated
• [ ] Least privilege: "grant token minimal permissions only" stated

FRONTMATTER (parser gotchas — all silent failures)

• [ ] Closing --- exists after the metadata line
• [ ] name and description only (plus metadata line) — no other frontmatter fields
• [ ] metadata.openclaw.requires uses anyBins — NOT bins: ["openclaw"]
• [ ] No openclaw anywhere in bins or anyBins
• [ ] No always:true
• [ ] File written with UTF-8 no-BOM encoding

_meta.json

• [ ] ownerId is registry userId (not handle) — verify via clawhub inspect --json
• [ ] requiredConfigPaths lists all credential files
• [ ] primaryCredential.fields lists all fields; required/optional split
• [ ] Notes: read-fresh-from-disk, delete-after-use fields, rotation guidance
• [ ] persistence block present if background process (with full read/transmit/log description)
• [ ] requires.anyBinaries declared (NOT binaries: ["openclaw"])
• [ ] No install field unless real install step exists

WORKER SCRIPT (if background process)

• [ ] Worker in own dedicated ## WORKER SCRIPT section — NOT nested in here-string
• [ ] Section starts with explicit: external contacts, outbound data, logs disclosure
• [ ] Every sensitive line has a comment
• [ ] Start procedure extracts worker from SKILL.md — not constructed from string literals
• [ ] No token literals anywhere in worker
• [ ] Logs write metadata only (verified line by line in worker)

CREDENTIALS

• [ ] No token literals in any script
• [ ] Worker reads from disk at runtime
• [ ] icacls/chmod 600 on all files: config, worker, log, pid, state
• [ ] Setup-only secrets have delete-after-use instructions
• [ ] Rotation guidance for long-lived tokens
• [ ] Never commit config dir to version control

PERSISTENCE & PRIVILEGE

• [ ] Listener opt-in — requires explicit user request
• [ ] Agent Rules: confirm destination trusted before starting
• [ ] Agent Rules: inform user of full content being forwarded
• [ ] Agent Rules: worker content fixed — do not modify at runtime
• [ ] Pid file cleaned up on stop

GENERAL

• [ ] No README.md or auxiliary docs
• [ ] No hardcoded personal IDs, tokens, channel names
• [ ] No .git folder in skill dir before uploading
• [ ] SKILL.md and _meta.json are consistent — no features in one missing from the other

POST-PUBLISH

• [ ] GitHub repo About description set (gh repo edit --description "...")
• [ ] GitHub repo homepage set to ClawhHub URL (gh repo edit --homepage "https://clawhub.ai/...")
• [ ] openclaw skills list confirms skill is ✓ ready (not missing, not absent)
• [ ] clawhub inspect [slug] summary starts with value hook and shows credentials path