On a work device? Switch to a personal phone and email before organizing.

Read the safety basics →

Unionize.software

Independent worker field guide

Free. Public. No account required.

Skills

OpenClaw Skill Bundle

How the shipped OpenClaw skill tells agents to use the project safely and consistently.

The repository includes an agent skill at skills/unionize-organizing/SKILL.md. The MCP server supplies the public material; the skill supplies the operating judgment for using it.

Quick start

  1. Clone and install:
git clone https://github.com/unionize-software/unionize.software.git
cd unionize.software
pnpm install
  1. Point your agent tool at the skill file:
  • skills/unionize-organizing/SKILL.md
  1. Run the MCP server (recommended alongside the skill):
pnpm mcp:serve

What the skill changes

The skill tells an agent when to search the guide corpus, when to use the pathfinder, how to respect jurisdiction and review metadata, and when the safe response is to narrow the next step or recommend qualified help.

Without those instructions, an agent can retrieve a relevant page and still use it badly—for example, by flattening a contractor-status question, treating a practice-based checklist as law, or escalating before workers have compared facts.

What it reinforces

  • no sabotage guidance
  • no company-device organizing
  • no pretending public guides are legal advice
  • preference for careful fact comparison, trusted coworkers, and lawful next steps
  • explicit separation between legal orientation, organizing practice, and case-specific advice
  • smaller, reversible next steps when worker status or retaliation risk is unclear

Use it with the corpus

The intended stack is simple:

  1. The skill identifies the safety and reasoning rules.
  2. The MCP catalog or CLI finds relevant public pages.
  3. The agent reads the guide metadata and body before responding.
  4. The response states important limits and routes the worker toward human help when the page cannot resolve the case.

The skill should not be copied into a system that also sends worker context to unreviewed third-party services without a separate privacy assessment. Good response rules do not compensate for an unsafe data path.

Maintenance

Treat the skill and the corpus as one interface. When page types, metadata, tool names, or safety boundaries change, update the skill and run the MCP checks before distributing it again.

The standard is the same as the public site: an agent should help a worker become more precise and less exposed. It should never manufacture certainty merely because it can produce fluent text.