Voice-first, dialogue-driven AI Skill creation — turn domain expertise into production-ready Agent Skills through Socratic conversation.
Built on OpenClaw with BigModel (GLM-5) as the default LLM.
| Mode | Output | Use Case |
|---|---|---|
| 🔧 Skill Mode | One SKILL.md + scripts + tests |
Personal automation, project-specific tools |
| 👥 Plugin Mode | Complete plugin: plugin.json + N skills + commands + MCP config |
Enterprise digital employees, team distributions |
| 🏢 Enterprise Mode (Plugin Studio) | Multi-tenant deployment serving an entire company | Companies that want every employee to self-serve their own AI digital coworker |
Plugin Studio is the deployment kit: install once on a company server, employees talk to it via Slack/Telegram/Web/voice → it either installs a ready-made plugin matching their role, or walks them through generating a new one. Admin approval gate, internal marketplace, audit trail. See enterprise/README.md.
Plugin Mode generates enterprise-ready digital employees — bundles of skills wrapped as Claude Code Plugins, including Slack/Notion/Salesforce MCP integration. Inspired by Anthropic's knowledge-work-plugins. 10 baseline role templates included (Sales Ops, Finance Analyst, Customer Support, Recruiting, ...).
Domain experts describe their repetitive tasks through natural conversation. The agent:
- Asks probing questions to extract tacit knowledge (IDEATION → REFINEMENT)
- Generates a complete Skill with proper structure and scripts (GENERATION)
- Creates behavioral test cases from the conversation examples (EVAL)
- Runs tests and reports results (TEST)
- Iterates fixes based on failures (ITERATE)
- Exports to multiple agent frameworks (EXPORT)
The expert never needs to know any syntax — they just talk.
Browse examples/commit-message-enhancer/ for a single-skill output (12 dialogue decisions → SKILL.md + 8 evals + 4 framework exports).
Browse examples/sales-ops-plugin/ for a complete Sales Ops Associate digital employee:
- 4 skills (prospect-research, call-prep, pipeline-review, battlecard-update)
- 1 command (
/sales-ops:weekly-report) - Slack + Notion MCP server config
- Validated and packaged via
package_plugin.py→sales-ops.zip+ marketplace snippet
Skill mode produces an end-to-end reproducible output from 12 dialogue decisions:
- Conversation transcript (15 turns)
- Generated SKILL.md with scripts + glossary
- Auto-generated eval suite (8 behavioral tests)
- 4 framework exports
The generated script even catches fake API keys planted in test diffs.
Plugin Mode ships with role templates (workspace/skills/plugin-creator/references/role-templates/) for instant scaffolding:
| Role | Plugin Output |
|---|---|
| Sales Ops Associate | sales-ops/ — 4 skills + weekly-report command + Slack/Notion |
| Finance Analyst | finance-analyst/ — 6 skills (journal entry, reconciliation, variance analysis, ...) |
| Customer Support Specialist | customer-support/ — 5 skills (triage, drafting, escalation, KB) |
| Recruiting Specialist | recruiting/ — 4 skills (JD drafting, screening, notes, offers) |
| Marketing / Product / Legal / Data / Search / Productivity | See enterprise-roles.md |
| Format | Output | Target |
|---|---|---|
| Claude Code | SKILL.md + scripts/ |
Claude Code |
| OpenClaw | SOUL.md + AGENTS.md + skills/ |
OpenClaw |
| Cursor | .cursorrules |
Cursor IDE |
| Hermes | agent.yaml + prompts/ |
Hermes Agent |
| Generic | PROMPT.md |
Any LLM with system prompt support |
- OpenClaw 2026.3+
- Python 3.10+
- An OpenAI-compatible LLM API key
# Clone
git clone https://github.com/xiaopingfeng/skills-creator.git
cd skills-creator
# Install Python deps
pip install -r requirements.txt
# Register as an OpenClaw agent
openclaw agents add skills-creator --workspace ./workspace
# Configure your LLM (example: BigModel)
export OPENAI_API_KEY=your-api-key
export OPENAI_BASE_URL=https://open.bigmodel.cn/api/paas/v4
# Start the gateway
openclaw gateway
# Chat via terminal
openclaw tuiThe skills and scripts work independently:
# Generate a skill from a session file
python3 workspace/skills/skill-generator/scripts/validate_skill.py ./my-skill/
# Export to Cursor Rules
python3 workspace/skills/skill-exporter/scripts/export_skill.py \
--format cursor --session session.json --skill-dir ./my-skill/ --output ./exports/Expert Voice/Text → OpenClaw Talk Mode
→ Brain (SOUL.md + AGENTS.md drive 7-phase dialogue)
→ skill-elicitor (question bank, summary templates)
→ skill-generator (SKILL.md generation + validation)
→ eval-generator (behavioral test cases)
→ skill-tester (automated test runner)
→ skill-exporter (multi-framework export)
→ Response to Expert
workspace/
├── SOUL.md # Agent personality
├── AGENTS.md # 7-phase workflow
├── TOOLS.md # Environment config
├── USER.md # Expert profile (auto-populated)
└── skills/
├── skill-elicitor/ # Dialogue engine
│ ├── SKILL.md
│ └── references/
│ ├── question-bank.md # CN/EN question corpus
│ └── summary-templates.md # Phase transition summaries
├── skill-generator/ # SKILL.md producer
│ ├── SKILL.md
│ ├── scripts/
│ │ ├── init_skill.sh # Directory scaffolding
│ │ └── validate_skill.py # Format validation
│ └── references/
│ ├── skill-anatomy.md # Writing guidelines
│ └── example-skills.md # Reference examples
├── eval-generator/ # Test case creator
│ ├── SKILL.md
│ └── scripts/
│ └── generate_evals.py
├── skill-tester/ # Test runner
│ ├── SKILL.md
│ └── scripts/
│ ├── run_eval.sh
│ ├── run_suite.sh
│ └── analyze_results.py
└── skill-exporter/ # Multi-framework export
├── SKILL.md
├── scripts/
│ └── export_skill.py
└── references/
└── export-formats.md
- One question per turn — optimized for voice; multi-question turns confuse speakers
- Verbal summaries at phase transitions — the expert confirms understanding before proceeding
- Behavioral tests, not string matching — skills produce varied output; behavior matters
- LLM-judged phase transitions — no hard-coded gates; the agent decides when to advance
- Framework-agnostic knowledge — session.json captures decisions independently of any target format
MIT