← all repositories
yizhiyanhua-ai/fireworks-tech-graph

Diagrams for people who'd rather write prose than click a GUI

A Claude Code skill that turns natural language into production-ready SVG and PNG technical diagrams, with built-in fluency in AI/Agent architecture patterns.

7.4k stars Python Coding Assistants
fireworks-tech-graph
Velocity · 7d
+127
★ / day
Trend
steady
star history

What it does

fireworks-tech-graph is a Claude Code skill that takes a plain-English (or Chinese) description of a system and emits polished SVG diagrams, exported as 1920px PNGs via cairosvg. It knows 14 UML diagram types and speaks the vocabulary of modern AI infrastructure — RAG pipelines, Mem0 memory tiers, multi-agent orchestration, tool-call loops — without you wrestling with Mermaid syntax or draw.io’s alignment guides.

The interesting bit

The “executable style system” is the quietly clever part. Seven template-driven styles (Flat Icon, Dark Terminal, Blueprint, Notion Clean, Glassmorphism, Claude Official, OpenAI Official) encode their visual rules directly into the generator — fonts, arrow semantics, swim-lane logic, brand-color palettes — so the model doesn’t freestyle the aesthetics. An eighth “Dark Luxury” style is AI-authored from a reference markdown file, which feels like a flex but also a proof of the broader pattern: turning vague creative workflows into constrained, repeatable systems.

Key highlights

  • Semantic vocabulary baked in: LLMs render as double-border rectangles, agents as hexagons, vector stores as ringed cylinders — no prompt engineering required.
  • Arrow grammar: color and dash pattern encode read, write, async, and loop semantics automatically.
  • 40+ product icons with brand colors (OpenAI, Anthropic, Pinecone, Kafka, PostgreSQL, etc.).
  • Renderer-agnostic SVG: pure inline, no external font fetches; works with cairosvg, rsvg-convert, or headless Chrome.
  • Installs as a Claude Code skill via npx skills add, or clone directly into ~/.claude/skills/.

Caveats

  • Requires one of three PNG renderers installed separately (cairosvg recommended; puppeteer pulls ~150MB of Chromium).
  • The README’s feature count for styles is inconsistent: badges and early sections claim 7 styles, while later sections and a second README paste claim 8 (adding Dark Luxury). The Dark Luxury style exists but is not template-driven, which may affect reproducibility.
  • No explicit mention of offline LLM inference; appears to rely on Claude Code’s model access, so usage assumes an active Claude Code environment.

Verdict

Worth a look if you write a lot of architecture docs, READMEs, or investor decks and hate maintaining diagrams by hand. Skip it if you need real-time collaborative editing or pixel-perfect manual control — draw.io still wins for fussy polish, and Mermaid for quick inline markdown.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.