A prompt factory that lives inside your AI agent

What it does

This is a “Skills-first” system for generating AI image prompts. You talk to Claude Code or OpenAI Codex CLI in plain Chinese or English; the project routes your request to one of twelve domain-specific skills (portrait, design, product, video, etc.), which assemble a prompt from a 1,246-element SQLite library and a 675-prompt community corpus. A Python engine underneath handles cross-domain queries and design-system assembly if you prefer to call it directly.

The interesting bit

The project treats prompt engineering as infrastructure rather than craft. It ships .claude/skills/ and .codex/skills/ directories that agent frameworks auto-discover, so the “installation” is mostly cloning the repo and letting your CLI tool ingest the skill definitions. The v2.0 engine added a cross-domain query layer that boosted claimed database utilization from 40.3% to 79.9% by combining multiple domain element sets on the fly.

Key highlights

• Dual-agent support: Native skill packs for both Claude Code and OpenAI Codex CLI.
• Three generation modes: Portrait (502 elements), Cross-Domain (995 elements), and Design (SQLite + 37 YAML color palettes for 200k+ claimed combinations).
• Structured data layer: Universal Elements Library in SQLite, prompt framework in YAML, and community source prompts categorized by creator and domain.
• Backwards compatibility: v1.0 Python API (IntelligentGenerator) still works unchanged alongside the v2.0 CrossDomainGenerator.
• Variable sampling: Parameterized elements with deduplication logic to avoid repetitive outputs.

Caveats

• Skills dependency: The intended workflow requires Claude Code CLI or Codex CLI; the standalone Python API is secondary and less documented.
• Metrics clarity: The “40.3% to 79.9%” utilization figure and “200,000+ combinations” are stated in the README without methodology or reproducible benchmarks.
• Language: Primary documentation and skill definitions are in Chinese; English support exists but may require inference.

Verdict

Worth a look if you generate prompts at volume inside Claude Code or Codex and want a structured, team-shareable skill library rather than ad-hoc prompting. Skip it if you just need a one-off prompt or prefer lightweight Python scripts without agent-framework coupling.