Install
python3 "$SKILL_ROOT/scripts/fireworks.py" versionWho it is for
If you want to describe an architecture or flow in natural language and get a geometry-checked SVG plus a high-resolution PNG, this Agent Skill is a better fit. The same skill works in Codex and Claude Code without changes. It can also produce a validated SVG-to-GIF semantic animation, and offline interactive HTML.
You already need to be able to call this skill in those coding assistants, and your machine needs python3 and must be able to find $SKILL_ROOT/scripts/fireworks.py. If you only want to draw by hand and do not plan to let an Agent write files, the benefit is limited. The page names Codex and Claude Code only; it does not name other clients. Also do not let a possibly older style count in the repository title override the body: 11 generator styles plus 1 AI writing style (Dark Luxury).
First try in 3 minutes
The page does not give a separate marketplace install command. Start with a read-only check: when the skill root is already ready, run
python3 "$SKILL_ROOT/scripts/fireworks.py" version
The success signal is that the command runs and prints version information. It does not write diagram files. If $SKILL_ROOT does not exist yet, stop: the available page text does not say how to install the skill into Codex or Claude Code, so do not invent an installer.
After version succeeds, do one controlled render. The official example is the user saying “Generate a Mem0 memory architecture diagram, dark style”. The skill classifies it as Memory Architecture Diagram, Style 2, generates an SVG with swimlanes, cylinders, and semantic arrows, then exports a 1920px PNG, and reports mem0-architecture.svg / mem0-architecture.png. Before writing files, have the Agent repeat those two paths, and generate only after you confirm. Do not start with GIF or HTML: the available page text mentions those artifacts but gives no matching subcommands.
Usage examples
Example 1: Produce an architecture diagram from natural language. Use the Mem0 example above to verify the whole chain: classify → SVG → 1920px PNG → report filenames. It does not document all visual parameters of Style 2, and it does not say where Mem0 domain knowledge comes from. So acceptance only checks: the classification is spoken, and those two filenames appear in the report. If the Agent starts changing your business code, it has already left the diagram task.
Example 2: Export PNG from an existing SVG. The official CLI is
python3 "$SKILL_ROOT/scripts/fireworks.py" export-png diagram.svg diagram.png --width 1920
This overwrites or creates diagram.png. First confirm the input SVG exists and the output path is writable. When you need to avoid cropped text, add "text_policy": "strict" to the configuration: it refuses visible truncation before writing, keeps the full source labels in the report, and explains how to recover; Chinese descriptions use the available second line. The available page text does not give a full JSON filename. So you can only confirm this field and the Chinese second line exist; do not invent other policy values.
Limits
Style counts follow the body: 11 generator styles, plus Dark Luxury. There are also four more engineering-oriented styles that give executable conventions for C4 review, cloud deploy, event streams, and reliability troubleshooting, plus deep AI/Agent domain patterns and all 14 UML diagram types. The page does not name the 14 UML types, and it does not name the four engineering styles. If the name is missing, do not call it out.
The copyable commands are only version and export-png. GIF and interactive HTML are written as capabilities, but there is no command line. Do not put a second installer beside them. Clients are written only as Codex and Claude Code.
The first stop is version, or one PNG whose path you already confirmed. Any batch rendering, any unknown field besides text_policy, or any GIF/HTML generation must first have the Agent explain the basis and the files it will write; you confirm, then it continues.