Logo Creator — Intelligent Logo Design Skill

Available Tools

• replicate skill — Use it to discover and run the right Replicate model for symbol generation, upscaling, or background removal.
• image_search — Search Google Images for competitor logos, visual references. Batch with |||.
• web_search — Research the brand, industry, competitors. Batch with |||.
• Read tool — CRITICAL. Use to view every generated/composed image. This is how you see and judge your own work.
• Bash — Run scripts:
  • scripts/compose_logo.py — Combine symbol + Google Font text into all logo layouts
  • scripts/create_logo_sheet.py — Build HTML contact sheet for visual comparison
  • scripts/remove_bg.py — Local background removal fallback (if the Replicate result produces artifacts)

Visual Critique Process

This is the most important part of the skill. After EVERY image operation (generation, composition, background removal), you MUST:

1. Read the image using the Read tool on the output file path
2. Describe what you see in 1-2 sentences (to yourself, not to the user)
3. Run the checklist against what you see
4. Decide: keep, regenerate, or adjust

After generating a symbol:

Read("logos/brand/round-1/logomark-concept-name.webp")
→ "I see a hexagonal shape with an arrow motif, centered on white. Clean lines, no text. Good."
→ KEEP

Read("logos/brand/round-1/logomark-abstract-wave.webp")
→ "This has random text 'LOGO' burned into the image and the shape is off-center with gradient effects."
→ REJECT — regenerate with stronger 'no text, no gradients' anchors

After removing a background:

Read("logos/brand/round-1/symbol-transparent.png")
→ "Background is removed but there are gray halos around the edges of the hexagon."
→ REJECT — re-run with local fallback: python3 scripts/remove_bg.py input.webp output.png

After composing:

Read("logos/brand/composed/brand-combo-horizontal.png")
→ "Symbol and text are well-balanced. Font loads correctly, spacing looks good. The symbol reads clearly at this size."
→ KEEP

Never skip the Read step. If you didn't look at it, you don't know if it's good.

Background Removal

Replicate background-removal models are good but not perfect — especially on logo symbols they can leave halos or eat into thin lines.

Primary method: load the replicate skill and use an appropriate background-removal model.

If the result has artifacts (gray halos, jagged edges, eaten details), use the local fallback:

pip install rembg pillow onnxruntime  # first time only
python3 skills/logo-creator/scripts/remove_bg.py \
  logos/<brand>/round-1/symbol.webp \
  logos/<brand>/round-1/symbol-transparent.png

After either method: Always Read the transparent PNG to check for artifacts. If both methods produce poor results, the compose_logo.py auto-crop can work with the original white-background image — it crops whitespace using pixel analysis instead of AI.

Output Structure

logos/<brand-name>/
  round-1/                    # AI-generated symbols
    logomark-*.webp
    logomark-*-transparent.png
  composed/                   # Programmatically composed logos
    <brand>-logomark.png
    <brand>-wordmark.png
    <brand>-wordmark-tagline.png
    <brand>-combo-horizontal.png
    <brand>-combo-vertical.png
    <brand>-combo-icon-right.png
    sheet.html
  final/                      # Approved logos (with transparent versions)

Workflow

Phase 1: Brand Discovery

Goal: Understand what this logo needs to communicate.

1. Extract from user message: brand name, what it does, any preferences.
2. If critical info is missing, ask — max 3 focused questions.
3. If user gave enough context, proceed without asking.
4. Research the landscape — batch search:
   • web_search: "[brand name]" company, [industry] brand identity trends 2025
   • image_search: [competitor1] logo, [competitor2] logo, [industry] logo design
5. Synthesize internally:
   • Brand personality (2-3 adjectives)
   • Visual directions to explore
   • Colors to try later (start B&W)
   • What to avoid (competitor similarities)
   • 2-3 concept directions with different metaphors/symbols
6. Pick a Google Font for the brand personality:
   • Space Grotesk — modern tech, geometric
   • Inter — clean, versatile, startup
   • Playfair Display — elegant, editorial, luxury
   • DM Sans — friendly, approachable, SaaS
   • Outfit — contemporary, balanced
   • Sora — futuristic, geometric
   • Libre Baskerville — traditional, trustworthy
   • Rubik — rounded, playful, consumer
   • Or any other Google Font that fits

Phase 2: Symbol Generation

Goal: Create the core symbol/icon.

1. Read references/prompt-patterns.md for prompt formulas and universal anchors.
2. Generate 4-6 symbol variations using the replicate skill:
   • Use the logomark prompt pattern from the reference
   • Vary the CONCEPT across generations (different metaphors), not just style
   • Start monochrome: Using only black (#000000). Monochrome design.
   • Output to logos/<brand-name>/round-1/
3. VIEW EVERY RESULT with Read. For each generated image:
   • Read("logos/<brand>/round-1/<filename>.webp")
   • Describe what you see: shape, composition, cleanliness, any artifacts
   • Run the self-critique checklist (see below)
   • If it fails 2+ criteria: regenerate with adjusted prompt. Note what went wrong and fix it in the next prompt.
   • If it passes: keep it and note what works well.
4. Remove backgrounds on the best 2-3 symbols:
   • Use the replicate skill with a background-removal model
   • Read the transparent PNG to check for halos/artifacts
   • If bad: re-run with scripts/remove_bg.py (local fallback)
5. Build contact sheet, open it, and give the user the path:

   python3 skills/logo-creator/scripts/create_logo_sheet.py \
     logos/<brand-name>/round-1/ \
     logos/<brand-name>/round-1/sheet.html \
     --title "<Brand> Symbols — Round 1" && \
   open logos/<brand-name>/round-1/sheet.html
   

   Tell the user: "The interactive contact sheet is at logos/<brand>/round-1/sheet.html — you can toggle light/dark backgrounds and zoom into each symbol."
6. Ask user: Which direction(s) do you like? What to change?

Phase 3: Composition

Goal: Combine approved symbol + real typography into complete logo layouts.

python3 skills/logo-creator/scripts/compose_logo.py \
  --brand "BrandName" \
  --symbol "logos/<brand>/round-1/symbol-transparent.png" \
  --output-dir "logos/<brand>/composed/" \
  --font "Space Grotesk" \
  --weight 700 \
  --color "#1a1a2e" \
  --tagline "Optional tagline" \
  --layouts all

Produces 6 layouts: logomark, wordmark, wordmark-tagline, combo-horizontal, combo-vertical, combo-icon-right.

After composing — READ every output:

• Read("logos/<brand>/composed/<brand>-combo-horizontal.png") — check symbol-text balance
• Read("logos/<brand>/composed/<brand>-combo-vertical.png") — check nothing clips
• Read("logos/<brand>/composed/<brand>-wordmark.png") — check font rendered correctly

If anything looks wrong (font didn't load, proportions off, text clipped), re-run with adjusted parameters.

Build contact sheet, open it, and tell the user:

python3 skills/logo-creator/scripts/create_logo_sheet.py \
  logos/<brand>/composed/ \
  logos/<brand>/composed/sheet.html \
  --title "<Brand> — Compositions" --cols 2 && \
open logos/<brand>/composed/sheet.html

Tell the user: "All compositions are at logos/<brand>/composed/sheet.html — the contact sheet lets you compare all layouts side by side with light/dark background toggles."

Phase 4: Iterate & Refine

Possible refinements:

• Different font — try another Google Font family
• Different weight — 400 for elegance, 800+ for boldness
• Color — --color "#2563eb" for brand color
• Letter spacing — -0.01em tighter, 0.05em more open
• Uppercase — --text-transform uppercase
• Different symbol — generate new ones if concept is wrong
• Dark mode — --bg "#1a1a2e" --color "#ffffff"

Each iteration = adjust parameters → re-run compose → Read to verify → rebuild contact sheet.

Max 3 rounds. If stuck, rethink the concept.

Phase 5: Finalize & Deliver

1. Transparent versions of approved compositions:
   • Use the replicate skill for final transparent exports
   • Read to verify clean edges
   • Fallback to scripts/remove_bg.py if needed
2. Dark-mode versions if useful
3. Build and open final contact sheet:

   python3 skills/logo-creator/scripts/create_logo_sheet.py \
     logos/<brand>/final/ logos/<brand>/final/sheet.html \
     --title "<Brand> — Final Logo Package" --cols 2 && \
   open logos/<brand>/final/sheet.html
   

4. Present to user with explicit paths:
   • List all files in logos/<brand>/final/ with what each one is (logomark, wordmark, combo, etc.)
   • Tell user: "The interactive contact sheet is at logos/<brand>/final/sheet.html"
   • Note the Google Font used so they can use it in their own materials
   • Suggest next steps (test at small sizes, create social media profiles, etc.)

compose_logo.py Reference

Also accepts a JSON config file as first positional argument.

Features: auto-crops symbol whitespace, loads Google Fonts live via Playwright, single browser instance, transparent PNG support.

Self-Critique Checklist

Run against what you SEE (via Read) after EVERY generated image. Discard and regenerate if 2+ fail:

• Centered? Design is centered in frame, not drifting
• Clean background? Solid white, no noise or gray patches
• Simple enough? Could be recognized at 32x32 pixels
• No unwanted text? No random letters, words, or watermarks
• Scalable? No fine detail that vanishes small
• Professional? Looks like a real product's logo, not clip art
• Unique? Doesn't obviously copy a well-known logo
• On-brand? Communicates the intended personality
• Clean edges? (after remove_bg) No halos, no eaten lines, no gray fringe

Rules

1. ALWAYS Read images after generating/composing. No blind trust. Look at your work.
2. ALWAYS give the user file paths and open contact sheets. After building a contact sheet, run open <path> AND tell the user the path. Never just say "here are your logos" without showing where they are.
3. Never present more than 6 options at once. Curate ruthlessly.
4. Never generate without universal anchors. See references/prompt-patterns.md.
5. Always start monochrome. Color comes in refinement rounds.
6. Always use compose_logo.py for text. Never ask AI to render text.
7. Always build contact sheets. Visual comparison is essential. Always open the HTML after building it.
8. Descriptive filenames. logomark-arrow-minimal.png not image_abc123.png.
9. Max 3 refinement rounds. If stuck, rethink the concept.
10. Verify remove_bg quality. Read the transparent PNG. If it has artifacts, use the local fallback script.