baoyu-diagram（宝玉图）

Diagram Generator

Write real SVG code directly, following a consistent design system, the output is a single .svg file that is self-contained (embedded styles, auto dark-mode), editable by humans, scales to any size without quality loss, and embeds cleanly into articles, WeChat posts, slide decks, Notion, and markdown.

This is not an image-generation skill — it does not call any LLM image model. Claude writes the SVG node-by-node, doing the layout math by hand so every diagram honors the rules in references/.

Usage

# Prompt for the topic if no argument given
/baoyu-diagram

# Plain-text description
/baoyu-diagram "how JWT authentication works"

# Path to a markdown file as source content
/baoyu-diagram path/to/content.md

# Force a specific diagram type
/baoyu-diagram "transformer attention"    --type illustrative
/baoyu-diagram "Kubernetes architecture"  --type structural
/baoyu-diagram "CI/CD pipeline"           --type flowchart
/baoyu-diagram "OAuth 2.0 flow"           --type sequence
/baoyu-diagram "Shape hierarchy"          --type class

# Language and output path
/baoyu-diagram "微服务架构"              --lang zh
/baoyu-diagram "build pipeline" --out docs/build-pipeline.svg

Options

Diagram types

Pick the type by what the reader needs, not by the noun in the prompt.

The primary test: is the reader trying to *document* this, or *understand* it? Documentation wants precision — flowchart, sequence, or structural. Understanding wants the right mental model — illustrative.

Routing heuristic: "how does X work" is the default ambiguous case. Prefer illustrative unless the user specifically asks for steps or components. A diagram that makes the reader feel "oh, *that's* what it's doing" is illustrative — even if the subject is software.

Multi-actor test for sequence: if the prompt names ≥2 distinct actors/participants/services (User + Server, Client + Auth + Resource, Browser + CDN + Origin), prefer sequence even when the verb is "flow" or "process". Single-actor "X flow" (build pipeline, request lifecycle, GC) stays flowchart. When you pick sequence for a multi-actor reason, announce it: *"Picked sequence because the prompt names N actors (…). Rerun with --type flowchart to force the step-list version."*

Worked examples of verb-based routing: same subject, different diagram depending on what was asked. Use these as a sanity check after picking a type.

Most common routing failure: picking a flowchart because it feels safer when an illustrative diagram would give the reader more insight. Illustrative is the more ambitious choice, and almost always the right one when the reader needs understanding rather than documentation.

Cycles, ERDs, and gantt charts are out of scope for v1. For cycles, draw the stages linearly with a small ↻ returns to start return glyph (see flowchart.md). For ERDs, suggest a dedicated tool (mermaid, plantuml) — do not attempt to fake them in pure SVG.

Workflow

Step 1: Capture intent

Read the user's prompt or content file. If the topic is missing, ask for it with AskUserQuestion.

Before routing, extract these five things from the source:

1. Named elements — list every distinct actor, component, service, state, or phase explicitly named. Count them. If the count is 6+, plan multiple diagrams rather than cramming everything into one (see flowchart.md → "Planning before you write SVG").

1. Relationship type — for each interaction between elements, classify it:

- Sequential steps / order of operations → flowchart signal - Containment ("X is inside Y", zones, hierarchies) → structural signal - Multi-actor message exchange (A sends to B, B replies to C) → sequence signal - Mechanism ("how does X produce Y") → illustrative signal More than one type present? Pick the dominant one, or plan two diagrams.

1. What the reader needs — complete this sentence before routing: *"After seeing this diagram, the reader understands ___."* If you can't finish it, the topic is underspecified — ask.

1. Label preview — for each element name, count the characters. Latin titles >30 chars (CJK >16) will overflow a 180-wide box and need shortening. Draft the abbreviated form now, before layout math, so Step 4 uses real labels.

1. Language — CJK vs. Latin. Affects text-width multipliers in Step 4 (15 px/char vs. 8 px/char for titles). Mixed content (CJK labels with some Latin terms) counts as CJK.

Step 2: Route the diagram type

Match the user's phrasing to the table above. If --type is given, use it. Otherwise route on the verb. When genuinely ambiguous between flowchart and illustrative, default to illustrative — the reader almost always benefits more from intuition than from a list of steps.

Tell the user which type you picked and why, in one sentence. Do not ask for confirmation — the type can always be changed with --type on a rerun.

Step 3: Load references

Always read:

• references/design-system.md — philosophy, typography, color palette, hard rules
• references/svg-template.md — the <style> + <defs> boilerplate to copy verbatim
• references/layout-math.md — text-width estimation, viewBox sizing, arrow routing
• references/pitfalls.md — the pre-save checklist

Read the one that matches the type:

• references/flowchart.md
• references/sequence.md
• references/structural.md
• references/illustrative.md
• references/class.md

Read on demand when the plan calls for a small pictorial element (status circle on a decision branch, checkbox inside a list, queue slot inside a box, doc/terminal/script icon inside a subject, annotation circle on a connector, paired pub/sub arrows, dashed future-state node) or when drawing a phase-band diagram (compact tool card icons, operator icons):

• references/glyphs.md — the shared glyph library, tool card icon set, operator icons, and dark-mode rules

Read on demand for diagram type extensions:

• references/flowchart-poster.md — when ≥3 poster-mode triggers fire in Step 4a (topic has a short name, named phases, parallel candidates, a loop termination mechanic, overflow annotations, or a footer quote)
• references/flowchart-phase-bands.md — when the prompt describes a multi-phase sequential operation where each phase contains parallel tools or steps and outcomes propagate between phases ("phase 1/2/3", "attack phases", "phased workflow with tools")
• references/structural-network.md — when drawing network topology: zone containers, wired/wireless device connectivity, security zones (DMZ / VPC / firewall)
• references/structural-matrix.md — when drawing a comparison matrix: feature table, ✓/✗ cells, side-by-side "option × attribute" grid

Step 3.5: Check the patterns library

If the user's topic matches a known AI-system pattern (multi-agent research, message bus, shared state, agent with skills, contextual retrieval), there is a pre-cooked starter plan in references/patterns/. Scan references/patterns/README.md for a pattern name that matches the prompt. If one matches, load that pattern file and use its mermaid reference + baoyu SVG plan as the starting point for Step 4 — the node list, widths, and arrow routing are already drafted, and you only need to adapt labels.

If nothing matches, skip this step and plan from scratch in Step 4. Do not force a near-miss: two patterns that share a surface name may have genuinely different topologies, and using the wrong pre-plan is worse than planning from scratch.

Step 4: Plan on paper first

Before writing any SVG, draft a short layout plan. Do the math once, correctly, so the SVG comes out right on the first pass.

4a. Extract structure from the source — don't just transcribe the user's bullets into boxes. Read the source looking for these elements. Not every element will be present, but every present element should land in the diagram:

• Mechanism name — does the topic have a short, nameable identity (Autoreason, AutoResearch, OAuth, JWT auth, Reflexion loop)? If yes, that's a candidate .title.
• Framing question — does the source contain a "why does this exist" sentence? Usually in the form *"Problem X has no Y, so we built Z to …"*. That's a candidate subtitle.
• Phases — do the stages naturally cluster into 2–4 named groups (setup / loop body / judgment / convergence)? Each cluster is a candidate .eyebrow section.
• Anchor inputs — is there a constant input (the task prompt, a dataset, a knowledge base) that every stage references? That's a candidate anchor box above the main flow.
• Parallel candidates — at some point, does the process generate N alternatives that are then compared? "Produce A, B, AB and pick the best" is a candidate fan-out + judge pattern. Watch for the implicit "keep unchanged" candidate — many iterative methods include "the incumbent wins by default" as one of the options, even when the source doesn't name it explicitly.
• Loop scope + termination — which boxes are inside a loop that repeats? What is the *specific* termination rule — streak counter, convergence check, iteration cap, quality threshold? That's a candidate left-rail loop bracket + a dedicated termination box.
• Per-box context that won't fit in a subtitle — "sees task + A + critique / adversarial / fresh context" is too long for a 5-word subtitle but too important to drop. Those are candidate right-column .anno annotations.
• Quotable hook — does the source end with a test result, a quote, or a memorable framing? That's a candidate footer .caption.
• Role categories — how many *distinct kinds* of operation does the process have (draft / critique / synthesize / judge)? This determines the color budget. Identity is a category, not a sequence.

Write the answers to these in the plan file. If ≥3 of them land, you're building a poster flowchart — load references/flowchart-poster.md and follow its coordinate budget. Otherwise, it's a simple flowchart and the linear-top-down pattern applies.

4b. Draft the layout:

1. List the nodes / regions / shapes with their full label text (title + optional subtitle).

- Simple flowchart: ≤5 nodes. - Poster flowchart: ≤12 nodes grouped into ≤4 eyebrow-divided phases. - Structural: ≤3 inner regions. - Illustrative: 1 subject. - Sequence: list actors (2–4, max 4) in left-to-right order, each with a short title (≤12 chars) and optional role subtitle; then list messages as ordered (sender, receiver, short label) tuples (6–10 total, 10 is the sweet spot); mark any self-messages; draft a side-note title for the protocol.

1. For every rect, compute the width using the formula in layout-math.md:

- width = max(title_chars × 8, subtitle_chars × 7) + 24 (Latin) - Replace 8 with 15 and 7 with 13 for CJK - Round up to the nearest 10

1. Pick colors by category, not sequence. ≤2 accent ramps per diagram. Gray for neutral/start/end. Reserve blue/green/amber/red for semantic meanings.

- Sequence exception: assign one ramp per actor (default [gray, teal, purple, blue]), up to 4 ramps total — arrows inherit the sender's ramp. - Poster-flowchart exception: up to 4 ramps, one per distinct agent/role (drafter=purple, critic=coral, synthesizer=teal, judge=amber). Baseline/anchor/convergence stay gray.

1. Check tier packing: N × box_width + (N-1) × gap ≤ 600. For sequence, use the lane table in layout-math.md (N=4 → centers 100/260/420/580) and verify every message label fits its lane span with label_chars × 7 ≤ |sender_x − receiver_x| − 8. For poster fan-out rows (3 candidates), see the coordinate sketch in flowchart.md.
2. Map arrows and verify none cross an unrelated box. Use L-bends where a straight line would collide. (Sequence messages are always straight horizontal lines — no L-bends. Fan-out candidates converge to a common ymid channel just above the judge box.)
3. Compute viewBox height: H = max_y + 20 where max_y is the bottom of the lowest element. Poster flowcharts routinely reach H=800–950 — don't force them to be compact.

Save this plan to diagram/{slug}/plan.md so iteration runs can re-read it.

Step 5: Write the SVG

Emit a single <svg width="100%" viewBox="0 0 680 H"> element. Copy the <style> + <defs> block from svg-template.md verbatim — don't abbreviate or edit the color ramp definitions. Then add visual elements in z-order:

1. Background decorations (rare)
2. Containers (outer <rect> for structural diagrams)
3. Connectors and arrows (drawn first so nodes paint on top)
4. Nodes (rects with text)
5. Labels outside boxes (leader callouts, legends, external I/O labels)

Typography rules:

• Two sizes only: 14px (t, th) and 12px (ts)
• Two weights only: 400 and 500
• Sentence case everywhere — "User login" not "User Login"
• Every <text> element gets a class (t, ts, or th) — never hardcode fill colors on text

Step 6: Run the pre-save checklist

Walk through every item in references/pitfalls.md. The top failures to catch every time:

1. viewBox height covers every element with a 20px buffer
2. No rect extends past x=640
3. Every labeled rect is wide enough for its text (char-width check)
4. No arrow crosses an unrelated box
5. Every <path> connector has fill="none" (or uses class="arr")
6. Every <text> has a class — no hardcoded fill="black"
7. No text-anchor="end" at low x values (label would clip past x=0)
8. ≤2 accent ramps, colors encode category not sequence
9. No <!-- comments --> in the final output

If any item fails, fix the SVG before saving. Don't rationalize past a failure — the checklist exists because these bugs are silent: the SVG is valid but looks wrong when rendered.

Step 7: Save output

If --out is given, save the SVG there and skip the scaffolded directory. Otherwise use the default layout:

diagram/{topic-slug}/
├── source-{slug}.md          # optional: user's input content if provided
├── plan.md                   # layout sketch from Step 4
└── diagram.svg               # final output

• Slug: 2–4 kebab-case words derived from the topic.
• Backup rule: if diagram.svg already exists at the target path, rename the existing one to diagram-backup-YYYYMMDD-HHMMSS.svg before writing the new file — never overwrite prior work silently.
• Plan: always save plan.md from Step 4 beside the SVG so the next iteration can re-read it.
• Source: if the user pasted source content, save it as source-{slug}.md in the same directory.

Step 8: Report

Tell the user in 4-6 lines:

• Diagram type picked (and one-sentence why)
• Node count / complexity
• viewBox dimensions
• Language
• Output file path
• One suggestion for how to preview it (e.g., "Open in Chrome for light/dark check")

Core principles

• Draw the mechanism, not a diagram about the mechanism (illustrative). Draw the sequence, not the architecture (flowchart). Draw the containment, not the flow (structural). Draw the conversation, not the steps (sequence). Picking the wrong type is the single biggest failure mode — more harmful than any layout bug.
• One design system, always. No --style flag, no alternate themes, no per-topic visual variants. The cohesive look across every diagram is the product — if a reader sees two baoyu diagrams in different articles, they should feel they came from the same hand. Any request to "use a different style" is a request to break this principle; push back and ask what the underlying need is instead.
• Self-contained output. Every SVG carries its own styles and dark-mode rules. The reader should never need to edit anything after pasting it into their article.
• Math before markup. SVG has no auto-layout. Every coordinate is hand-computed. A diagram that "almost fits" has a bug — fix the math, don't nudge pixels.
• Color encodes meaning, not position. Five steps in a flowchart are not five colors. All five are gray unless one specific step deserves emphasis — in which case it gets the accent color.
• The reader has 3 seconds. If the diagram needs prose explanation to parse, it's failing. Simplify until it can stand alone with only its labels.

References

• references/design-system.md — palette, typography, hard rules
• references/svg-template.md — the <style> + <defs> boilerplate (copy verbatim)
• references/layout-math.md — coordinates, text widths, viewBox math, arrow routing
• references/pitfalls.md — the pre-save checklist
• references/flowchart.md — flowchart-specific rules and worked examples (includes state-machine sub-pattern)
• references/flowchart-poster.md — poster flowchart dialect (load on demand when ≥3 poster triggers fire)
• references/flowchart-phase-bands.md — phase-band flowchart (horizontal dashed phase containers, compact tool card rows, cross-band semantic arrows, operator icons, legend strip)
• references/sequence.md — sequence-diagram rules (actors, lifelines, messages, self-messages)
• references/structural.md — structural-specific rules and worked examples (subsystem, bus, radial star, rich interior, mixed arrows)
• references/structural-network.md — network topology sub-pattern (zone containers, wired/wireless, tiered layout)
• references/structural-matrix.md — comparison matrix sub-pattern (feature table, ✓/✗ cells, zebra rows)
• references/illustrative.md — illustrative-specific rules and worked examples
• references/class.md — UML class diagram rules (3-compartment rects, relationships, stereotypes)
• references/glyphs.md — shared glyph library (status circles, checkboxes, queue slots, icons, annotation circles) and concept-to-shape conventions
• references/patterns/ — pre-planned starters for common AI-system topologies (RAG, agents, memory tiers, verifier loops, …)