Logo Svg Generator

Generate multi-size SVG logos from natural-language product briefs, following Simple Icons design rules; runs an iterative generate → visual-check → optimize loop with a vision-capable LLM. Invoke when the user asks to design a logo, brand mark, app icon or product icon from a description.

Turn a natural-language product brief (name, purpose, audience, mood, colors) into a production-ready set of monochrome / brand-color SVG logos in multiple sizes (16, 24, 32, 64, 128, 256, 512 px + 24×24 Simple-Icons-compatible master), by following the design regulations mined from the Simple Icons project.

The skill uses an iterative visual-review loop: generate → rasterize → visually inspect (vision LLM) → critique → optimize → re-inspect → finalize.

Portable / agent-agnostic. This skill is a plain Markdown + Python bundle. It runs anywhere a coding agent or LLM tool can load a SKILL.md and call local scripts — including but not limited to TRAE, Opencode, Claude Code, Hermes, openclaw, Cursor, Continue, or a custom harness. Any reference to "the agent" / "the host" below applies uniformly to all of them.

---

When to invoke

Trigger this skill whenever the user says any of:

• "Design / make / create a logo for …"
• "I need an SVG icon / brand mark / product logo for …"
• "Generate a logo for my product …"
• "帮我设计 / 生成一个 logo / 图标"
• "给我做一个 SVG 的品牌标识"

Do NOT invoke for: free-form illustration, complex marketing artwork, photo editing, or non-vector deliverables.

---

Inputs the skill collects (or infers)

Before generation, gather (asking the user only when critical, otherwise infer sensibly):

Field	Example
product_name	"NovaSync"
product_purpose	"cross-device file sync tool"
audience / vibe	"developers, minimal, techy"
preferred_color (hex)	#4F46E5 (if omitted → pick from mood)
mono_or_color	monochrome + branded variant (default: both)
sizes	[16, 24, 32, 64, 128, 256, 512] (default)
metaphor_hints	"letter N + sync arrows" (optional)

---

Core workflow (MUST follow in order)

Step 1 — Read the design regulations

Load references/design-rules.md — the condensed Simple Icons style guide (viewBox 0 0 24 24, single <path>, no stroke, no fill attribute in master, touches ≥ 2 sides of viewBox, centered, single-color-with-0.5px-gap technique, slug rules).

Also skim references/motif-patterns.md for common shape recipes (roundel, letterform, monogram, arrow, waveform, chevron, gear, chat-bubble, spark, etc.) with real path-fragment examples.

Step 2 — Brainstorm 2–3 concept directions

Given the brief, write a short internal design brief covering:

1. Metaphor (what visual idea encapsulates the product?)
2. Shape family (geometric / letterform / abstract / pictogram)
3. Color (primary hex + when mono is required)
4. Composition (single symbol vs. mark + wordmark — this skill outputs symbol only, no wordmark)

Pick the strongest concept unless the user explicitly wants alternatives.

Step 3 — Generate the 24×24 master SVG

Follow these hard rules (from Simple Icons CONTRIBUTING.md):

• Root: <svg role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg">
• One <title> with the product name
• Exactly one <path d="…"/>
• No <rect> <circle> <ellipse> <polygon> <g> <line> <defs> <style> <use> <mask> <clipPath>
• No fill / stroke / class / id on the path (color is added later per variant)
• Path must touch at least 2 sides of the viewBox (no padding)
• Visual center around (12, 12); keep coordinates to ≤ 3 decimals
• Where color layers would exist, use a 0.5-unit gap trick to imply separation while staying single-path

Write the master to output/<slug>/master-24.svg.

Step 4 — Rasterize + visual check (VISION-CAPABLE LLM REQUIRED BY DEFAULT)

Hard requirement: the logo check step MUST be performed by an LLM that has image-input (vision) capability, and that vision capability MUST actually be exercised on the rendered PNG. This is the default and non-optional path — text-only critique is not acceptable as a substitute. If the current host / model does not expose vision input, the skill MUST NOT silently fall back; instead the agent MUST:

1. Explicitly inform the user that a vision-capable LLM is required for the review loop.
2. Ask the user to switch to a vision-capable model (or provide an explicit --no-review / structural-only override) before continuing.
3. Only proceed to the structural-only degradation (Step 5, last bullet) after the user has knowingly opted in.

Run scripts/render_previews.py to rasterize the master at 32, 64, 128, 256, 512 and produce a review sheet PNG (output/<slug>/_review.png) with:

• The icon on white, on black, and inside a rounded-square app-icon frame
• The 5 sizes side-by-side (to catch scale-dependent issues)
• Coordinate grid overlay at 64px

Then, hand _review.png plus the source SVG to the vision-capable LLM with the prompt in assets/templates/visual-review-prompt.md. The image MUST be attached as a real image input (not described in text) so the model exercises its vision pathway. The reviewer must return a JSON critique:

json

{
  "score": 0-10,
  "readable_at_16px": true,
  "issues": [
    {"severity": "high|med|low", "area": "balance|proportion|weight|negative-space|readability|brand-fit|geometry", "note": "…"}
  ],
  "concrete_fixes": ["Nudge the top-left arc 0.3u down", "Increase counter of letter N by 8% ..."],
  "verdict": "accept|revise"
}

Step 5 — Optimize

If verdict == "revise", apply the concrete_fixes to the SVG path (edit coordinates directly; keep single-path rule). Save as master-24.v{N}.svg, then loop back to Step 4.

• Max loop count: 3 iterations (config in scripts/config.json; can be raised to 5 with --max-iterations when the caller explicitly asks for more polish).
• Stop early when score ≥ 8.5 AND readable_at_16px == true AND no high severity issue remains.
• If loops exhaust without acceptance, keep the highest-scoring version as the final master, save every iteration under output/<slug>/history/ for debugging, and produce a user-facing summary (output/<slug>/_review-summary.md) that translates each residual issue.note into a plain-language sentence plus a suggested next action (e.g. "the letter N is slightly off-center — you can nudge it 0.3u right, or ask me to run 2 more iterations").
• Graceful degradation when no vision-capable LLM is available: this path is opt-in only and requires either the user's explicit consent (see Step 4 hard requirement) or a --no-review / require_vision_llm=false override in scripts/config.json. When permitted, skip Step 4/5's vision critique and instead run the deterministic checklist in scripts/validate_svg.py plus the geometric checks documented in references/design-rules.md §2 (touches-two-sides, bbox center window, ink coverage). Mark brand.json.review_mode = "structural-only" so downstream users know a human should eyeball the result. Never engage this path silently.

Step 6 — Generate multi-size + color variants

Once the master is accepted, run scripts/export_variants.py to produce:

text

output/<slug>/
├─ master-24.svg              (canonical Simple-Icons-compatible)
├─ mono/
│  ├─ logo-16.svg  logo-24.svg  logo-32.svg  logo-64.svg
│  ├─ logo-128.svg logo-256.svg logo-512.svg
├─ color/
│  └─ (same set, path filled with brand hex)
├─ png/
│  └─ logo-16.png … logo-512.png  (PNG previews)
├─ favicon/
│  └─ favicon.ico  apple-touch-icon-180.png
├─ _review.png
└─ brand.json                 ({ slug, title, hex, source_brief, iterations })

Multi-size SVGs are the same path with adjusted width/height attributes and (for small sizes 16 & 24) optional path-simplification (see scripts/simplify_for_small.py) — remove sub-details that vanish below 20px.

Step 7 — Deliver

Give the user computer:// links to the top-level output folder, the master SVG, and the review sheet. Summarize the concept in ≤ 3 sentences.

---

Non-negotiable design rules (quick check before finalizing)

Every SVG file the skill emits must pass all of:

• role="img", viewBox="0 0 24 24", xmlns — in that order
• Exactly one <title>
• Exactly one <path>, no other tags
• Master has no fill, no stroke, no class, no id
• Path touches ≥ 2 sides of the 24×24 canvas
• Visually centered (bounding box center within (11, 13) × (11, 13))
• Aspect ratio of original concept preserved (never stretched)
• Slug follows slugify rules (lowercase alphanumeric; .→dot, +→plus, &→and, remove spaces / punctuation)
• At 16 px it is still recognisable (validated by vision LLM)

---

File layout of this skill project

text

logo-svg-generator-skill/
├─ SKILL.md                    ← this file (SkillHub entry point)
├─ references/
│  ├─ design-rules.md          ← condensed Simple Icons rules
│  ├─ motif-patterns.md        ← common shape recipes + path snippets
│  └─ sample-icons.md          ← 15 curated real SVG references
├─ assets/
│  ├─ logo.svg / logo-*.svg    ← this project's own brand marks
│  └─ templates/
│     ├─ visual-review-prompt.md  ← prompt for vision LLM critique
│     └─ concept-brief-template.md
├─ scripts/
│  ├─ fetch_samples.py         ← pull N icons from simple-icons CDN
│  ├─ render_previews.py       ← SVG → review PNG sheet
│  ├─ export_variants.py       ← multi-size + color variants
│  ├─ simplify_for_small.py    ← small-size path simplification
│  ├─ validate_svg.py          ← rule compliance checker
│  ├─ deploy_skill.py          ← copy skill files → target agent's skill dir
│  └─ config.json
├─ samples/                    ← downloaded reference SVGs (git-ignored)
└─ output/                     ← per-run deliverables (git-ignored)

---

Related references

• references/design-rules.md
• references/motif-patterns.md
• references/sample-icons.md
• assets/templates/visual-review-prompt.md
• assets/templates/concept-brief-template.md

Source of truth for design regulations:

• https://simpleicons.org
• https://github.com/simple-icons/simple-icons/blob/develop/CONTRIBUTING.md
• https://github.com/simple-icons/simple-icons/blob/develop/slugs.md

---

Edge cases & how to handle them

Situation	What the skill does
Very long product name (>16 chars, e.g. "HyperContinuousDeliveryPlatform")	Do NOT try to fit the whole name inside the 24×24 canvas. Use only the first initial as a monogram, or ask the user to pick a 1–2-letter abbreviation. Full name still goes into <title> and brand.json.title.
Multi-word name ("Nova Sync", "Pulse Mail")	Compose either an initials monogram (NS, PM) or a symbol derived from the dominant word. Never render the full wordmark inside the icon.
Non-ASCII product name ("音见 AI", "Café")	Slugify to ASCII per §5 of design-rules.md before saving files; keep the original characters only inside <title>.
User requests a wordmark or lockup	Explain that this skill emits symbol marks only, then either (a) generate the strongest single-letter monogram, or (b) hand off to a general design step outside the skill.
User requests a photorealistic / gradient / 3-D logo	Refuse politely and explain that Simple-Icons style is monochrome flat vector. Offer a flat interpretation as an alternative.
Brand color has poor contrast on white (e.g. #F7F7F7)	Warn the user, still emit the color variant, and always ship a black monochrome fallback.
Vision LLM unavailable / rate-limited	Stop and inform the user — a vision-capable LLM is the default, non-optional reviewer. Only fall back to review_mode = "structural-only" (see Step 5) after the user explicitly opts in (or --no-review is passed). Do NOT block the export once the user has opted in.
Rasterizer unavailable (no cairosvg, resvg, or inkscape)	Emit the SVG variants + brand.json, skip png/ + _review.png, and tell the user which one-liner install would restore raster output.
CDN blocked / offline	fetch_samples.py retries with the fallback mirrors in scripts/config.json. If all fail, skip sample fetching — the skill can still design from references/motif-patterns.md alone.
User in mainland China	The default primary CDN is jsDelivr; alternate mirrors (fastly, unpkg, bytedance jsdelivr mirror) are tried in order. See scripts/config.json.cdn_url_templates.

---

FAQ / anti-patterns

Q: The reviewer keeps returning score = 6-7 — what should I do? The most common cause is too many motifs. Cut back to one dominant idea (§Combining motifs · Rule 1 in motif-patterns.md) and re-run. If the score still stalls, ask the user for a fresh metaphor hint rather than looping indefinitely.

Q: validate_svg.py fails with element: <g> is forbidden — why? An LLM sometimes wraps the path in <g> for grouping. Delete the <g> tags; keep only <svg><title/><path/></svg>. The skill MUST emit exactly two child elements of <svg>.

Q: The path uses fill="currentColor" on the master. That is still a fill attribute and fails compliance. The master carries NO fill; color is added by export_variants.py only for the color/ variants.

Q: My icon looks great at 512 px but is unreadable at 16 px. Run scripts/simplify_for_small.py and, more importantly, drop sub-details below ~1u wide before Step 3. The reviewer's readable_at_16px field is a hard gate; if it comes back false, do NOT accept.

Q: Can I skip the vision-review loop for speed? Yes — set max_review_iterations: 0 in scripts/config.json or pass --no-review. You will get a rule-compliant but visually un-audited SVG. Recommended only when the caller is doing their own review.

Q: The generated logo looks generic (a circle with a letter). Feed the concept brief more specific metaphor_hints. "Two orbiting nodes with a break", "envelope with an EKG waveform" produces markedly stronger results than "modern and clean".

Anti-patterns to avoid

1. Silently swallowing rule violations. Every emitted SVG must pass validate_svg.py. If it fails, fix the SVG or report the violation to the user — never emit anyway.
2. Adding a wordmark inside the 24×24 canvas. Even a 2-letter wordmark rarely survives at 16 px.
3. Using <circle> "because the geometry is easier". Convert to a <path> arc; multiple SVG elements are a hard-fail.
4. Skipping the concept brief. Without a brief, the reviewer has no brand_fit reference and the iteration loop cannot converge.
5. Looping past max_review_iterations. If the score has not converged in 3 rounds, further iterations rarely help — ask the user for a new metaphor instead.

---

End-to-end walk-through

text

User: 帮我给 NovaSync 设计一个 logo，跨设备文件同步工具，靛蓝色，开发者向。

Step 1  → skill loads references/design-rules.md and references/motif-patterns.md
Step 2  → concept brief (auto-derived):
          product_name: "NovaSync", slug: "novasync",
          preferred_color: "#4F46E5",
          metaphor_hints: "letter N + sync orbit arrows",
          shape_family: "letterform"
Step 3  → generates master-24.svg with a single <path> that renders an 'N'
          whose diagonal is a partial orbit arc; touches top + bottom edges.
Step 4  → python scripts/render_previews.py output/novasync/master-24.svg
          → output/novasync/_review.png
          vision LLM returns:
            { "score": 7.4, "readable_at_16px": true,
              "issues": [ {"severity":"high","area":"balance",
                           "note":"orbit arc is 0.4u too low"} ],
              "concrete_fixes": ["shift arc anchor from (7,15) to (7,14.6)"],
              "verdict": "revise" }
Step 5  → path is edited; saved as master-24.v1.svg; loop back.
          v1 scores 8.7 → verdict "accept".
Step 6  → python scripts/export_variants.py output/novasync/master-24.v1.svg \
              --slug novasync --title NovaSync --hex "#4F46E5"
Step 7  → user receives:
            output/novasync/
              master-24.svg
              mono/logo-{16,24,32,64,128,256,512}.svg
              color/logo-{...}.svg      (fill="#4F46E5")
              png/logo-{...}.png
              _review.png
              brand.json                { "review_mode": "vision", ... }

For a copy-pasteable input brief see examples/01-novasync/brief.md; expected outputs are described in examples/01-novasync/README.md.