🦞   ENTERPRISE AI TOOLKIT Series · Article 01
Article I — Skills & Plugins for Enterprise AI

Skills
Creator

A Socratic interlocutor that turns the things experts know how to do but cannot quite write down — into production-ready Agent Skills.

Built on OpenClaw Powered by GLM-5 MIT · 2026
§ I — Thesis

The expert
need not type.

Most agent skills today are written by engineers who have never done the work. Skills Creator inverts that: an agent interrogates the practitioner, draft after draft, until the tacit knowledge has been pinned to the page — as a SKILL.md, an eval suite, a runnable test harness.

The expert speaks. The agent listens. Twelve to fifteen turns later, a complete, tested skill exists where one did not before — and exports cleanly to whichever framework the expert's team happens to use today, or tomorrow.

The skill is not the artifact.
The dialogue is.
— DESIGN PRINCIPLE No. 1
§ II — Specimen Transcript

From silence to shipped in fifteen turns.

An excerpt from a real session that produced commit-message-enhancer, included verbatim in the repository as a reproducible reference.

AgentWhat repetitive task do you wish Claude could handle for you?
ExpertWriting commit messages. Every day. I forget the type, or write too much.
AgentWalk me through a recent example.
ExpertYesterday I added JWT refresh tokens. Wrote: feat(auth): add JWT refresh token support.
⌐ eleven turns later — 6 IDEATION, 6 REFINEMENT, 1 confirmation
Agent8 of 8 evals passed. The script even caught a fake API key planted in a test diff. Ship it?
ExpertShip it.
· · · · ·
→ 17 files generated · 4 framework exports
§ III — Method

Seven phases, each
under five minutes.

Transitions are LLM-judged on conversation quality, not hard-coded gates. Phases collapse or repeat as the dialogue demands.

I
Ideation
Two concrete examples of the task. The trigger. The output. Who consumes it.
II
Refinement
Edge cases. Failure modes. Quality criteria. Domain vocabulary. Security.
III
Generation
A complete skill emerges: SKILL.md, scripts, references. Validated against schema.
IV
Eval
Five to eight behavioral test cases auto-built from the dialogue's examples.
V
Test
The skill runs against every eval. Failures reported in plain language.
VI
Iterate
Each failure becomes a proposed amendment. Loops back to Test until clean.
VII
Export
Same dialogue, four frameworks: Claude Code, OpenClaw, Cursor, Hermes, Generic.
§ IV — Exhibits — Output Formats

One conversation,
five destinations.

The decisions captured in session.json are framework-agnostic. Adapters render them into the dialect of each target.

Exhibit A

Claude Code

Anthropic · Default
SKILL.md
scripts/*
references/*
Exhibit B

OpenClaw

Persistent Daemon
SOUL.md
AGENTS.md
TOOLS.md
skills/*
Exhibit C

Cursor

IDE Rules
.cursorrules
Exhibit D

Hermes

Nous Research
agent.yaml
prompts/system.md
prompts/examples/*
Exhibit E

Generic

Any system prompt
PROMPT.md
§ V — Case File

A complete
reproducible specimen.

Twelve dialogue decisions become one Claude Code skill, eight eval cases, and four framework exports — seventeen files total, all browsable below.

§ V·b — Plugin Mode — Enterprise Digital Employee

From skill to colleague.

The same Socratic method scales up: a single dialogue can produce not just a skill, but a complete Claude Code Plugin — a digital employee bundling 4–7 skills, slash commands, and MCP server connections (Slack, Notion, Salesforce, ...). Modeled on Anthropic's knowledge-work-plugins.

Role A

Sales Ops

4 skills · Slack + Notion
prospect-research
call-prep
pipeline-review
battlecard-update
/weekly-report
Role B

Finance Analyst

6 skills · local files
journal-entry
reconciliation
variance-analysis
financial-statements
close-management
audit-support
Role C

Customer Support

5 skills · Slack
ticket-triage
response-drafting
escalation-routing
kb-article-generation
canned-responses
Role D

Recruiting

4 skills · local
jd-drafting
resume-screening
interview-notes-synthesis
offer-letter-generation
+ 6 More

Marketing · Product · Legal · Data · Search · Productivity

10 baseline templates
See enterprise-roles.md
§ V·c — Plugin Studio — Enterprise Deployment

Install once. Every employee self-serves.

One server. The whole company talks to it. Sales reps get their sales-ops plugin pre-built. New finance hires self-install finance-analyst. Teams without a template walk the agent through 30 minutes of dialogue and ship a custom plugin with admin sign-off.

Component A

install.sh

One-command setup
bash install.sh \
  --target-dir /opt/plugin-studio
Component B

employee-intake

Role detection skill
match → install
no match → create
Component C

plugin-publisher

Marketplace + approval
validate → review
→ publish → notify
Component D

enterprise.config.json

Per-company tuning
company · admins
marketplace · roles
§ VI — Installation Directives

Run it on
your machine.

# 1. Acquire the dossier $ git clone https://github.com/fxp/skills-creator.git $ cd skills-creator && pip install -r requirements.txt # 2. Register as an OpenClaw agent $ openclaw agents add skills-creator --workspace ./workspace # 3. Provision an LLM (any OpenAI-compatible endpoint) $ export OPENAI_API_KEY=your-key $ export OPENAI_BASE_URL=https://your-api/v1 # 4. Begin the interview $ openclaw gateway && openclaw tui
APPROVED
★ View on GitHub ★