PolicyForge Guidelines

PolicyForge is a conversational prototype for converting general agent policies into target harness permission drafts. In Phase 1, the goal is not full automation. The goal is to validate demand, identify semantic mismatch patterns, and collect high-quality conversion cases for future deterministic tooling.

Included

• CLAUDE.md Claude-oriented project instructions for running the prototype in Claude Code.
• CURSOR.md Cursor-oriented project instructions for running the same workflow.
• README.zh.md Chinese overview of the submission.
• EXAMPLES.md Example conversion inputs, outputs, and testing prompts.
• .claude-plugin/ Placeholder directory for future Claude-oriented packaging.
• .cursor/rules/ Cursor rules for the project prompt workflow.
• skills/ Core PolicyForge conversion skill prompt.

Phase 1 Scope

• Input: APoP-style JSON and simplified generic JSON policy files
• Output: a constrained Claude Code permissions draft
• Method: prompt-based semantic mapping with explicit downgrade behavior
• Safety model: forced manual review before any real use

Important Limitation

This submission is a prompt prototype, not a production-safe compiler. Any generated permission result must be manually reviewed.

Suggested Usage

1. Open this folder in Claude Code or Cursor.
2. Load the corresponding instruction file.
3. Ask the agent to convert a policy JSON file into Claude Code permissions.
4. Review the output for structure, semantic fidelity, and downgrade handling.