name: fireworks-tech-graph description: >- Create technical diagrams such as software architecture, data flow, flowcharts, sequence diagrams, agent/memory systems, UML, ER, network topology, timelines, and technical concept maps, then export SVG+PNG. Use when the user asks to draw or visualize a system, workflow, protocol, model, or engineering concept. Do not use for photos, raster illustrations, image editing, decorative artwork, or quantitative data charts.

Fireworks Tech Graph

Generate production-quality SVG technical diagrams exported as PNG via cairosvg (recommended), rsvg-convert, or puppeteer.

Runtime Compatibility

Use this repository unchanged in both Codex and Claude Code. It follows the Agent Skills layout: SKILL.md is the shared entry point, bundled resources use relative paths, and agents/openai.yaml adds optional Codex UI metadata without affecting Claude Code.

Before reading a reference or running a script, resolve the directory containing this SKILL.md as SKILL_ROOT. Do not assume the current working directory is the skill directory, and do not assume a variable set in one shell call persists into the next.

• In Claude Code, use ${CLAUDE_SKILL_DIR}.
• In Codex, use the absolute skill directory shown in the loaded skill metadata.

Every command block below sets SKILL_ROOT itself. In Codex, replace /absolute/path/from-codex-skill-metadata with the absolute skill directory before running the command.

Helper Scripts (Recommended)

Four helper scripts in scripts/ directory provide stable SVG generation and validation:

1. generate-diagram.sh - Validate SVG + export PNG

SKILL_ROOT="${CLAUDE_SKILL_DIR:-/absolute/path/from-codex-skill-metadata}"
"$SKILL_ROOT/scripts/generate-diagram.sh" -t architecture -s 1 -o ./output/arch.svg

• Validates an existing SVG file
• Exports PNG after validation
• Example: "$SKILL_ROOT/scripts/generate-diagram.sh" -t architecture -s 1 -o ./output/arch.svg

2. generate-from-template.py - Create starter SVG from template

SKILL_ROOT="${CLAUDE_SKILL_DIR:-/absolute/path/from-codex-skill-metadata}"
mkdir -p ./output
python3 "$SKILL_ROOT/scripts/generate-from-template.py" architecture ./output/arch.svg '{"title":"My Diagram","nodes":[],"arrows":[]}'

• Loads a built-in SVG template
• Renders nodes, arrows, and legend entries from JSON input
• Escapes text content to keep output XML-valid

3. validate-svg.sh - Validate SVG syntax

SKILL_ROOT="${CLAUDE_SKILL_DIR:-/absolute/path/from-codex-skill-metadata}"
"$SKILL_ROOT/scripts/validate-svg.sh" <svg-file>

• Checks XML syntax
• Verifies tag balance
• Validates marker references
• Checks attribute completeness
• Validates path data

4. test-all-styles.sh - Batch test all styles

SKILL_ROOT="${CLAUDE_SKILL_DIR:-/absolute/path/from-codex-skill-metadata}"
"$SKILL_ROOT/scripts/test-all-styles.sh"

• Tests multiple diagram sizes
• Validates all generated SVGs
• Generates test report

When to use scripts:

• Use scripts when generating complex SVGs to avoid syntax errors
• Scripts provide automatic validation and error reporting
• Recommended for production diagrams

When to generate SVG directly:

• Simple diagrams with few elements
• Quick prototypes
• When you need full control over SVG structure

Workflow (Always Follow This Order)

1. Classify the diagram type (see Diagram Types below)
2. Extract structure — identify layers, nodes, edges, flows, and semantic groups from user description
3. Plan layout — apply the layout rules for the diagram type
4. Load style reference — always load $SKILL_ROOT/references/style-1-flat-icon.md unless user specifies another; load the matching $SKILL_ROOT/references/style-N-*.md for exact color tokens and SVG patterns
5. Map nodes to shapes — use Shape Vocabulary below
6. Check icon needs — load $SKILL_ROOT/references/icons.md for known products
7. Write SVG with adaptive strategy (see SVG Generation Strategy below)
8. Validate: Run "$SKILL_ROOT/scripts/validate-svg.sh" file.svg to check XML, marker references, arrow-component collisions, and renderability
9. Export PNG: Use cairosvg (recommended). Load $SKILL_ROOT/references/png-export.md when choosing another renderer
10. Report the generated file paths
11. Visual review gate — if your runtime can read images, load the exported PNG back and inspect it. Syntactic validity does not guarantee visual correctness: arrows may cross through component interiors, labels may collide with lifelines or other labels, boxes may overlap, alt-frame text may sit on top of a message, or a legend may cover content. If you see any of these, revise the SVG and re-export, with at most two focused correction passes. Common fixes:
    • Route arrows through gaps between boxes, not through box interiors
    • Move arrow labels 6-8px away from the arrow line (offset-first); add background rects only when offset is insufficient
    • Widen inter-row/inter-column gutters so same-layer arrows have clear corridors
    • Collapse repeated cross-layer arrows into a single "delegates down" rail outside the content area
    • Move legend/notes out of any region where arrows or labels land
    • Increase viewBox height/width rather than packing elements tighter
    • If a filtered element (drop-shadow, blur) is missing one side of its border, move it ≥30px away from that viewBox edge, or remove the filter and rely on color/contrast for visual separation Report visual_review: passed after inspection. If image reading is unavailable, report visual_review: skipped (image reader unavailable) — do not guess or claim visual correctness.

Rule Precedence

Use this order when instructions disagree:

1. The user's explicit content and style request
2. The selected $SKILL_ROOT/references/style-N-*.md visual tokens (palette, typography, corner radius, shadow treatment)
3. Diagram-type layout rules and semantic flow requirements in this file
4. Universal defaults and examples

Geometry and validation gates always remain active: style guidance cannot justify unreadable text, missing marker definitions, or arrows crossing component interiors. Tables in this file define semantic defaults; a selected style may override their colors and stroke treatment while preserving the meaning and direction of each flow.

Diagram Types & Layout Rules

Architecture Diagram

Nodes = services/components. Group into horizontal layers (top→bottom or left→right).

• Typical layers: Client → Gateway/LB → Services → Data/Storage
• Use <rect> dashed containers to group related services in the same layer
• Arrow direction follows data/request flow
• ViewBox: 0 0 960 600 standard, 0 0 960 800 for tall stacks

Data Flow Diagram

Emphasizes what data moves where. Focus on data transformation.

• Label every arrow with the data type (e.g., "embeddings", "query", "context")
• Use wider arrows (stroke-width: 2.5) for primary data paths
• Dashed arrows for control/trigger flows
• Color arrows by data category (not just Agent/RAG — use semantics)

Flowchart / Process Flow

Sequential decision/process steps.

• Top-to-bottom preferred; left-to-right for wide flows
• Diamond shapes for decisions, rounded rects for processes, parallelograms for I/O
• Keep node labels short (≤3 words); put detail in sub-labels
• Align nodes on a grid: x positions snap to 120px intervals, y to 80px

Agent Architecture Diagram

Shows how an AI agent reasons, uses tools, and manages memory. Key conceptual layers to always consider:

• Input layer: User, query, trigger
• Agent core: LLM, reasoning loop, planner
• Memory layer: Short-term (context window), Long-term (vector/graph DB), Episodic
• Tool layer: Tool calls, APIs, search, code execution
• Output layer: Response, action, side-effects Use cyclic arrows (loop arcs) to show iterative reasoning. Separate memory types visually.

Memory Architecture Diagram (Mem0, MemGPT-style)

Specialized agent diagram focused on memory operations.

• Show memory write path and read path separately (different arrow colors)
• Memory tiers: Working Memory → Short-term → Long-term → External Store
• Label memory operations: store(), retrieve(), forget(), consolidate()
• Use stacked rects or layered cylinders for storage tiers

Sequence Diagram

Time-ordered message exchanges between participants.

• Participants as vertical lifelines (top labels + vertical dashed lines)
• Messages as horizontal arrows between lifelines, top-to-bottom time order
• Activation boxes (thin filled rects on lifeline) show active processing
• Group with <rect> loop/alt frames with label in top-left corner
• ViewBox height = 80 + (num_messages × 50)

Comparison / Feature Matrix

Side-by-side comparison of approaches, systems, or components.

• Column headers = systems, row headers = attributes
• Row height: 40px; column width: min 120px; header row height: 50px
• Checked cell: tinted background (e.g. #dcfce7) + ✓ checkmark; unsupported: #f9fafb fill
• Alternating row fills (#f9fafb / #ffffff) for readability
• Max readable columns: 5; beyond that, split into two diagrams

Timeline / Gantt

Horizontal time axis showing durations, phases, and milestones.

• X-axis = time (weeks/months/quarters); Y-axis = items/tasks/phases
• Bars: rounded rects, colored by category, labeled inside or beside
• Milestone markers: diamond or filled circle at specific x position with label above
• ViewBox: 0 0 960 400 typical; wider for many time periods: 0 0 1200 400

Mind Map / Concept Map

Radial layout from central concept.

• Central node at cx=480, cy=280
• First-level branches: evenly distributed around center (360/N degrees)
• Second-level branches: branch off first-level at 30-45° offset
• Use curved <path> with cubic bezier for branches, not straight lines

Class Diagram (UML)

Static structure showing classes, attributes, methods, and relationships.

• Class box: 3-compartment rect (name / attributes / methods), min width 160px
  • Top compartment: class name, bold, centered (abstract = italic)
  • Middle: attributes with visibility (+ public, - private, # protected)
  • Bottom: method signatures, same visibility notation
• Relationships:
  • Inheritance (extends): solid line + hollow triangle arrowhead, child → parent
  • Implementation (interface): dashed line + hollow triangle, class → interface
  • Association: solid line + open arrowhead, label with multiplicity (1, 0.., 1..)
  • Aggregation: solid line + hollow diamond on container side
  • Composition: solid line + filled diamond on container side
  • Dependency: dashed line + open arrowhead
• Interface: <<interface>> stereotype above name, or circle/lollipop notation
• Enum: compartment rect with <<enumeration>> stereotype, values in bottom
• Layout: parent classes top, children below; interfaces to the left/right of implementors
• ViewBox: 0 0 960 600 standard; 0 0 960 800 for deep hierarchies

Use Case Diagram (UML)

System functionality from user perspective.

• Actor: stick figure (circle head + body line) placed outside system boundary
  • Label below figure, 13-14px
  • Primary actors on left, secondary/supporting on right
• Use case: ellipse with label centered inside, min 140×60px
  • Keep names verb phrases: "Create Order", "Process Payment"
• System boundary: large rect with dashed border + system name in top-left
• Relationships:
  • Include: dashed arrow <<include>> from base to included use case
  • Extend: dashed arrow <<extend>> from extension to base use case
  • Generalization: solid line + hollow triangle (specialized → general)
• Layout: system boundary centered, actors outside, use cases inside
• ViewBox: 0 0 960 600 standard

State Machine Diagram (UML)

Lifecycle states and transitions of an entity.

• State: rounded rect with state name, min 120×50px
  • Internal activities: small text entry/ action, exit/ action, do/ activity
  • Initial state: filled black circle (r=8), one outgoing arrow
  • Final state: filled circle (r=8) inside hollow circle (r=12)
  • Choice: small hollow diamond, guard labels on outgoing arrows [condition]
• Transition: arrow with optional label event [guard] / action
  • Guard conditions in square brackets
  • Actions after /
• Composite/nested state: larger rect containing sub-states, with name tab
• Fork/join: thick horizontal or vertical black bar (synchronization)
• Layout: initial state top-left, final state bottom-right, flow top-to-bottom
• ViewBox: 0 0 960 600 standard

ER Diagram (Entity-Relationship)

Database schema and data relationships.

• Entity: rect with entity name in header (bold), attributes below
  • Primary key attribute: underlined
  • Foreign key: italic or marked with (FK)
  • Min width: 160px; attribute font-size: 12px
• Relationship: diamond shape on connecting line
  • Label inside diamond: "has", "belongs to", "enrolls in"
  • Cardinality labels near entity: 1, N, 0..1, 0..*, 1..*
• Weak entity: double-bordered rect with double diamond relationship
• Associative entity: diamond + rect hybrid (rect with diamond inside)
• Line style: solid for identifying relationships, dashed for non-identifying
• Layout: entities in 2-3 rows, relationships between related entities
• ViewBox: 0 0 960 600 standard; wider 0 0 1200 600 for many entities

Network Topology

Physical or logical network infrastructure.

• Devices: icon-like rects or rounded rects
  • Router: circle with cross arrows
  • Switch: rect with arrow grid
  • Server: stacked rect (rack icon)
  • Firewall: brick-pattern rect or shield shape
  • Load Balancer: horizontal split rect with arrows
  • Cloud: cloud path (overlapping arcs)
• Connections: lines between device centers
  • Ethernet/wired: solid line, label bandwidth
  • Wireless: dashed line with WiFi symbol
  • VPN: dashed line with lock icon
• Subnets/Zones: dashed rect containers with zone label (DMZ, Internal, External)
• Labels: device hostname + IP below, 12-13px
• Layout: tiered top-to-bottom (Internet → Edge → Core → Access → Endpoints)
• ViewBox: 0 0 960 600 standard

UML Coverage Map

Full mapping of UML 14 diagram types to supported diagram types:

Shape Vocabulary

Map semantic concepts to consistent shapes across all diagram types:

Arrow Semantics

Always assign arrow meaning, not just color. The values below are defaults; the selected style reference overrides colors and stroke weights while preserving flow semantics:

Always include a legend when 2+ arrow types are used.

Layout Rules & Validation

Spacing:

• Same-layer nodes: 80px horizontal, 120px vertical between layers
• Canvas margins: 40px minimum, 60px between node edges
• Snap to 8px grid: horizontal 120px intervals, vertical 120px intervals

Arrow Labels (CRITICAL):

• Offset-first (default): place label 6-8px above horizontal arrows, or 8px left/right of vertical arrows — do not overlap the arrow line
• Background fallback: add <rect fill="canvas_bg" opacity="0.95"/> only when the offset label still crosses another visual element (another arrow, a node edge, etc.)
• Place mid-arrow, ≤3 words, stagger by 15-20px when multiple arrows converge
• Maintain 10px safety distance from nodes

Arrow Routing:

• Prefer orthogonal (L-shaped) paths to minimize crossings
• Anchor arrows on component edges, not geometric centers
• Route around dense node clusters, use different y-offsets for parallel arrows
• Jump-over arcs (5px radius) for unavoidable crossings
• Compress equivalent bidirectional traffic only when both directions share the same semantics and styling: use one corridor with marker-start + marker-end, or two visibly offset paths in that corridor
• Keep read/write, request/response, sync/async, or differently labeled directions as separate arrows; remove redundant bends and duplicate rails without erasing direction or meaning

Post-Generation Arrow Optimization:

When a user asks to "优化箭头" / "fix arrow routing" / "optimize the diagram" on an already-generated diagram, preserve all nodes, containers, styles, and layout — only modify the arrows entries in the JSON data, then re-render with generate-from-template.py.

Available arrow override fields (in recommended order of use):

For JSON/template rendering, the default remains "badge" for backward compatibility. Set "label_style": "offset" on individual arrows when you want offset-first labels without background rects.

Optimization steps:

1. Read the existing SVG — identify which arrows overlap, cross nodes, or look misaligned
2. Find those arrows in the JSON data by source / target pair
3. Add source_port / target_port if the exit/entry direction is wrong; add corridor_x / corridor_y to space parallel arrows apart; use route_points only when hints alone cannot resolve the path
4. Re-run generate-from-template.py with the updated JSON and validate with validate-svg.sh

Example — spacing two overlapping arrows into separate corridors:

{ "source": "nodeA", "target": "nodeB", "corridor_y": [280] }
{ "source": "nodeC", "target": "nodeD", "corridor_y": [320] }

Line Overlap Prevention (CRITICAL - common in AI-generated diagrams): When two arrows must cross each other, ALWAYS use jump-over arcs to prevent visual overlap:

• Crossing horizontal arrows: add a small semicircle arc (radius 5px, stroke same color as arrow, fill none) that "jumps over" the other line
• SVG pattern for jump-over: use a white/matching-background arc on the lower layer, then draw the upper arc on top
• Multiple crossings: stagger arc radii (5px, 7px, 9px) so arcs don't overlap each other
• Never let two arrows' straight-line segments cross without a jump-over arc

Validation Checklist (run before finalizing):

1. Arrow-Component Collision: Arrows MUST NOT pass through component interiors (route around with orthogonal paths)
2. Text Overflow: All text MUST fit with 8px padding (estimate: text.length × 7px ≤ shape_width - 16px)
3. Arrow-Text Alignment: Arrow endpoints MUST connect to shape edges (not floating); arrow labels should not overlap arrow lines (use offset positioning or background rects)
4. Container Discipline: Prefer arrows entering and leaving section containers through open gaps between components, not through inner component bodies
5. Filter Boundary Safety: For every element with filter="url(...)", verify (element_x + element_width + filter_extension) ≤ viewBox_width AND element_x ≥ filter_extension. The default filter region extends 10-20% beyond bbox; staying near viewBox edges causes Chrome/cairosvg to clip the element's edge-side stroke (one side of the border vanishes while other sides render correctly)
6. Arrow-Title Collision: Arrows MUST NOT cross through section/container title text or region labels (font-size ≥ 13px). For smaller annotations (< 13px), prefer routing around but tolerate if layout constraints require it. (Visual self-review check — not covered by validate-svg.sh automated checks)
7. Frame Label–Arrow Alignment (sequence diagrams): Section/frame label badges MUST be vertically centered with their first message arrow. Compute badge_y = first_arrow_y - (badge_height / 2). When appending new sections to an existing diagram, verify alignment matches the existing sections — this is the most common regression when adding content incrementally. Use variables in Python list generation to enforce the constraint: sec_y = 840; badge_y = sec_y - 9 # for height=18 badge
8. Marker Integrity: Every marker-start, marker-mid, and marker-end URL MUST resolve to a <marker id="..."> definition
9. Visual Review Status: Report whether the exported PNG was visually inspected; automated validation does not cover every text, legend, or arrow-arrow collision

SVG Technical Rules

• ViewBox: 0 0 960 600 default; 0 0 960 800 tall; 0 0 1200 600 wide
• Fonts: embed via <style>font-family: ...</style> — no external @import (cairosvg / rsvg-convert cannot fetch external URLs)
• <defs>: arrow markers, gradients, filters, clip paths
• Text: minimum 12px, prefer 13-14px labels, 11px sub-labels, 16-18px titles
• All arrows: <marker> with markerEnd, sized markerWidth="10" markerHeight="7"
• Drop shadows: <feDropShadow> in <filter>, apply sparingly (key nodes only)
• Curved paths: use M x1,y1 C cx1,cy1 cx2,cy2 x2,y2 cubic bezier for loops/feedback arrows
• Clip content: use <clipPath> if text might overflow a node box
• Z-order (drawing order): SVG uses painter's model — later elements cover earlier ones. Recommended layer order (bottom → top): ① canvas background ② dashed containers / region backgrounds ③ arrows and connection lines ④ node shapes (rects, circles) ⑤ text labels and annotations ⑥ legends and overlays. When arrows pass near text, draw arrows BEFORE text so text stays readable. Adjust per diagram needs — this is guidance, not rigid.

SVG Generation & Error Prevention

MANDATORY: Python List Method (ALWAYS use this):

python3 << 'EOF'
lines = []
lines.append('<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 960 700">')
lines.append('  <defs>')
# ... each line separately
lines.append('</svg>')

with open('/path/to/output.svg', 'w') as f:
    f.write('\n'.join(lines))
print("SVG generated successfully")
EOF

Why mandatory: Prevents character truncation, typos, and syntax errors. Each line is independent and easy to verify.

Pre-Tool-Call Checklist (CRITICAL - use EVERY time):

1. ✅ Can I write out the COMPLETE command/content right now?
2. ✅ Do I have ALL required parameters ready?
3. ✅ Have I checked for syntax errors in my prepared content?

If ANY answer is NO: STOP. Do NOT call the tool. Prepare the content first.

Error Recovery Protocol:

• First error: Analyze root cause, apply targeted fix
• Second error: Switch method entirely (Python list → chunked generation)
• Third error: STOP and report to user - do NOT loop endlessly
• Never: Retry the same failing command or call tools with empty parameters

Validation (run after generation):

python3 -c "import xml.etree.ElementTree as ET; ET.parse('file.svg')" && echo "✓ Valid XML"
# Or use cairosvg as a render-time check:
python3 -c "import cairosvg; cairosvg.svg2png(url='file.svg', write_to='/tmp/test.png')" && echo "✓ Renders" && rm /tmp/test.png

If using generate-from-template.py:

• Prefer source / target node ids in arrow JSON so the generator can snap to node edges
• Keep x1,y1,x2,y2 as hints or fallback coordinates, not the main routing primitive
• Let the generator choose orthogonal routes; avoid hardcoding center-to-center straight lines unless the path is guaranteed clear

Common Syntax Errors to Avoid:

• ❌ yt-anchor → ✅ y="60" text-anchor="middle"
• ❌ x="390 (missing y) → ✅ x="390" y="250"
• ❌ fill=#fff → ✅ fill="#ffffff"
• ❌ marker-end= → ✅ marker-end="url(#arrow)"
• ❌ L 29450 → ✅ L 290,220
• ❌ Missing </svg> at end
• ❌ Element with filter near viewBox edge — filter region extends 20% (default) or more beyond bbox; if that region exceeds viewBox, Chrome/cairosvg clip the filter rendering AND can drop the element's own stroke on that side. Keep filtered elements at least max(20% of element size, shadow blur radius × 3) away from viewBox edges, or omit the filter.

Output

• Default: ./[derived-name].svg and ./[derived-name].png in current directory
• Custom: user specifies path with --output /path/ or 输出到 /path/
• PNG export: see SVG → PNG Conversion below

SVG → PNG Conversion

Use $SKILL_ROOT/scripts/generate-diagram.sh by default. Load $SKILL_ROOT/references/png-export.md only when selecting a renderer manually, handling CJK/emoji fallback, converting browser-generated SVG, or using the bundled Puppeteer converter.

Styles

Load the matching $SKILL_ROOT/references/style-N-*.md for exact color tokens and SVG patterns.

Style Selection

Default: Style 1 (Flat Icon) for most diagrams. Load $SKILL_ROOT/references/style-diagram-matrix.md for detailed style-to-diagram-type recommendations.

These patterns appear frequently — internalize them:

RAG Pipeline: Query → Embed → VectorSearch → Retrieve → Augment → LLM → Response Agentic RAG: adds Agent loop with Tool use between Query and LLM Agentic Search: Query → Planner → [Search Tool / Calculator / Code] → Synthesizer → Response Mem0 / Memory Layer: Input → Memory Manager → [Write: VectorDB + GraphDB] / [Read: Retrieve+Rank] → Context Agent Memory Types: Sensory (raw input) → Working (context window) → Episodic (past interactions) → Semantic (facts) → Procedural (skills) Multi-Agent: Orchestrator → [SubAgent A / SubAgent B / SubAgent C] → Aggregator → Output Tool Call Flow: LLM → Tool Selector → Tool Execution → Result Parser → LLM (loop)

English | 中文

fireworks-tech-graph

Stop drawing diagrams by hand. Describe your system in English or Chinese — get publication-ready SVG + PNG technical diagrams in seconds.

Overview

fireworks-tech-graph is one Agent Skill that works unchanged in Codex and Claude Code. It turns natural language descriptions into polished SVG diagrams, then exports them as high-resolution PNG via cairosvg (recommended), with rsvg-convert and puppeteer available as alternatives. It ships with 7 template styles and 1 AI-authored style (Dark Luxury) and deep knowledge of AI/Agent domain patterns (RAG, Agentic Search, Mem0, Multi-Agent, Tool Call flows), plus full support for all 14 UML diagram types.

User: "Generate a Mem0 memory architecture diagram, dark style"
  → Skill classifies: Memory Architecture Diagram, Style 2
  → Generates SVG with swim lanes, cylinders, semantic arrows
  → Exports 1920px PNG
  → Reports: mem0-architecture.svg / mem0-architecture.png

Work With the Builder

This project is also a proof surface for a broader capability: turning vague AI/devtool workflows into constrained, reusable systems with validation, documentation, export paths, and product-facing polish.

If you are building agent infrastructure, AI IDEs, internal copilots, developer tools, technical documentation systems, or applied AI workflow products, I am open to scoped paid sprints, design-partner work, and founding engineer conversations.

• Founder-facing profile: https://bradzhang.dev/en
• Commercial case study: https://bradzhang.dev/en/case-studies/fireworks-tech-graph
• Work with me: https://bradzhang.dev/en/work-with-me

Showcase

All samples are exported at 1920px width (2× retina) by the regression pipeline. It prefers cairosvg and falls back to rsvg-convert. PNG keeps technical text and line work lossless.

Style 1 — Flat Icon (default)

Mem0 Memory Architecture — white background, semantic arrows, layered memory system

Style 2 — Dark Terminal

Tool Call Flow — dark background, neon accents, monospace font

Style 3 — Blueprint

Microservices Architecture — deep blue background, grid lines, cyan strokes

Style 4 — Notion Clean

Agent Memory Types — minimal white, single accent color

Style 5 — Glassmorphism

Multi-Agent Collaboration — dark gradient background, frosted glass cards

Style 6 — Claude Official

System Architecture — warm cream background (#f8f6f3), Anthropic brand colors, clean professional aesthetic

Style 7 — OpenAI Official

API Integration Flow — pure white background, OpenAI brand palette, modern minimalist design

Style 8 — Dark Luxury (AI-authored)

Agent Runtime Architecture — control plane, execution and state layers, champagne-gold structure, semantic color buckets

Stable Prompt Recipes

Use prompts like these when you want the model to stay close to the repo's strongest regression-tested outputs:

Style 1 — Flat Icon

Draw a Mem0 memory architecture diagram in style 1 (Flat Icon).
Use four horizontal sections: Input Layer, Memory Manager, Storage Layer, Output / Retrieval.
Include User, AI App / Agent, LLM, mem0 Client, Memory Manager, Vector Store, Graph DB, Key-Value Store, History Store, Context Builder, Ranked Results, Personalized Response.
Use semantic arrows for read, write, control, and data flow. Keep the layout clean and product-doc friendly.

Style 2 — Dark Terminal

Draw a tool call flow diagram in style 2 (Dark Terminal).
Show User query, Retrieve chunks, Generate answer, Knowledge base, Agent, Terminal, Source documents, and Grounded answer.
Use terminal chrome, neon accents, monospace typography, and semantic arrows for retrieval, synthesis, and embedding update.

Style 3 — Blueprint

Draw a microservices architecture diagram in style 3 (Blueprint).
Create numbered engineering sections like 01 // EDGE, 02 // APPLICATION SERVICES, 03 // DATA + EVENT INFRA, 04 // OBSERVABILITY.
Include Client Apps, API Gateway, Auth / Policy, three services, Event Router, Postgres, Redis Cache, Warehouse, and Metrics / Traces.
Use blueprint grid, cyan strokes, and a bottom-right title block.

Style 4 — Notion Clean

Draw an agent memory types diagram in style 4 (Notion Clean).
Compare Sensory Memory, Working Memory, Episodic Memory, Semantic Memory, and Procedural Memory around a central Agent core.
Use a minimal white layout, neutral borders, one accent color for arrows, and short storage tags for each memory type.

Style 5 — Glassmorphism

Draw a multi-agent collaboration diagram in style 5 (Glassmorphism).
Use three sections: Mission Control, Specialist Agents, and Synthesis.
Include User brief, Coordinator Agent, Research Agent, Coding Agent, Review Agent, Shared Memory, Synthesis Engine, and Final response.
Use frosted cards, soft glow, and semantic arrows for delegation, shared memory writes, and synthesis output.

Style 6 — Claude Official

Draw a system architecture diagram in style 6 (Claude Official).
Use left-side layer labels: Interface Layer, Core Layer, Foundation Layer.
Include Client Surface, Gateway, Task Planner, Model Runtime, Policy Guardrails, Memory Store, Tool Runtime, Observability, and Registry.
Use warm cream background, restrained brand-like palette, generous whitespace, and a bottom-right legend.

Style 7 — OpenAI Official

Draw an API integration flow diagram in style 7 (OpenAI Official).
Use three sections: Entry, Model + Tools, and Delivery.
Include Application, OpenAI SDK Layer, Prompt Builder, Model Runtime, Tool Calls, Response Formatter, Observability, and Release Control.
Keep the look minimal, white, precise, and modern with clean green-accented arrows.

Style 8 — Dark Luxury (AI-authored)

Style 8 is not a template-driven style. The AI reads references/style-8-dark-luxury.md and hand-crafts the SVG directly.

Draw an Agent Runtime Architecture diagram in style 8 (Dark Luxury).
Use two sections: Control Plane and Execution and State.
Include Client, Gateway, Agent Runtime, Vector Memory, Tool Runtime, and Trace + Eval.
Use a deep black background (#0a0a0a), champagne gold (#d4a574) for titles and cluster labels,
and spread node colors across the full color wheel: emerald, violet, sky blue, rose, amber, cool-gray.
Apply Georgia serif only for the main title and section labels (≥11px); use sans-serif for all node text and arrow labels.

Features

• 8 visual styles — 7 template-driven (Flat Icon to OpenAI Official) + 1 AI-authored (Dark Luxury)
• Executable style system — style guides are encoded into the generator, not only documented in markdown
• 14 diagram types — Full UML support (Class, Component, Deployment, Package, Composite Structure, Object, Use Case, Activity, State Machine, Sequence, Communication, Timing, Interaction Overview, ER Diagram) plus AI/Agent domain diagrams
• AI/Agent domain patterns — RAG, Agentic Search, Mem0, Multi-Agent, Tool Call, and more built-in
• Semantic shape vocabulary — LLM = double-border rect, Agent = hexagon, Vector Store = ringed cylinder
• Semantic arrow system — color + dash pattern encode meaning (write vs read vs async vs loop)
• Structured SVG validation — XML parsing, marker-start/mid/end integrity, and arrow-component collision checks for M/L/H/V/Q/C/S/T paths
• Visual review gate — exported PNGs are inspected for clipping, overlap, label placement, and routing regressions before delivery
• Product icons — 40+ products with brand colors: OpenAI, Anthropic, Pinecone, Weaviate, Kafka, PostgreSQL…
• Swim lane grouping — automatic layer labeling for complex architectures
• SVG + PNG output — SVG for editing, 1920px PNG for embedding
• Renderer-friendly — pure inline SVG, no external font fetching; renders cleanly in cairosvg, rsvg-convert, and headless Chrome

Loop Engineering

The first render is treated as a candidate, not an automatic final result. fireworks-tech-graph uses an agent-driven, bounded validation feedback loop to move each diagram toward a verified deliverable:

Prompt
  → Diagram Contract
  → Semantic IR
  → Style Spec
  → Route Planner
  → SVG Build
  → Structural Validation
  → PNG Visual Readback
  → Targeted Revision
  → Verified SVG + PNG

The loop follows five design principles:

1. Evaluate, don't assert — completion is backed by validator and render evidence, not by the model saying the diagram looks correct.
2. Deterministic checks first — XML structure, marker integrity, path geometry, arrow-component collisions, and renderability are checked before visual judgment.
3. Perceptual validation second — the exported PNG is read back to inspect clipping, label collisions, hierarchy, whitespace,