Logo as code skill

Vectorizing hand-drawn logos

Overview

A hand-drawn logo is charming but unmeasurable: it lives as a flat image, or as an SVG of hundreds of hand-placed polyline points — smooth to the eye, but rough, impossible to edit, and not truly scalable. This skill turns each contour into a small set of smooth cubic Bézier curves — genuine parametric polynomials

B(t) = (1-t)³P₀ + 3(1-t)²t·C₁ + 3(1-t)t²·C₂ + t³P₃,   t ∈ [0,1]

The curve is fitted along the original outline (not replaced by idealized circles/rectangles), so the hand-drawn fullness and character — even an off-centre curl or a wobble you want to keep — survive. The output is fully mathematical: reproducible, infinitely scalable, parametrically tweakable, and diff-able in git. From one definition you fan out color/layout variants and rasterize to PNG/ICO.

Core insight: resample the outline to even spacing → lightly smooth out hand jitter → fit a closed Catmull-Rom spline → emit it as exact cubic Béziers. The fit interpolates every sample point and is everywhere C¹ (tangent-continuous), which is the mathematical source of the smoothness.

When to use

• "Turn my hand-drawn / sketched logo into an SVG / into code."
• "Make my logo reproducible, scalable, editable; generate favicon / app-icon sizes."
• "Recreate this mark precisely as math / parametric curves; give me color variants."
• Any mascot, wordmark, monogram, or brand mark that should become a versioned asset.

Not for: generating a brand-new logo design from scratch (this reconstructs an existing drawing); photographic images; logos that are already clean vector files and only need recoloring (just edit the SVG directly).

The toolkit in this skill

• scripts/vectorize.py — the generic engine (stdlib only). Importable library of the fit pipeline + SVG emission. You do not edit this per logo.
• scripts/build_gallery.py — generic, stdlib-only review-gallery generator: builds a self-contained gallery.html (original vs reconstruction comparison + variant matrix).
• scripts/rasterize.js — generic SVG → PNG/ICO/favicon exporter (sharp + png-to-ico).
• scripts/generate_showcase.py — optional AI showcase renders (Gemini / Nano Banana). The only part that needs Python deps + an API key; everything else is dependency-free.
• assets/gallery_template.html — the gallery template build_gallery.py fills.
• examples/chan-monkey/build.py — a complete worked example; the template to copy.
• references/ — the math, tuning, layer-mapping, bitmap input, quality checklist, showcase styles.

Workflow

Follow these steps. Keep the user in the loop at the visual comparison step (4) — that is where the craft is.

1. Get a vector polyline SVG of the drawing

The engine needs an SVG whose paths use absolute M/L/H/V/C/Z.

• Already a vector/polyline SVG (e.g. exported from a drawing app)? Proceed.
• A bitmap (scan/photo/PNG)? Trace it to SVG first — see references/bitmap-input.md (potrace / autotrace, then flatten to absolute paths).

2. Inspect the SVG and map its contours to roles

Read the SVG's <path d="…"> data. Each M…Z is one closed contour. Identify what each one is: outer silhouette, interior holes (a face, a counter in a letter), features painted on top, wordmark glyphs, and any contour that is really a circle (eyes, dots). See references/layer-structure.md for how holes, the even-odd "knockout" trick, and circles work.

3. Scaffold a logo project from the template

Create a new repo/folder for the logo and copy in the two scripts. Write a build.py modeled on examples/chan-monkey/build.py that declares, for this logo:

• the contour→role mapping (step 2),
• per-contour TUNING = (resample_step_px, smoothing_passes),
• how roles compose into even-odd paths / circles (the render function),
• the VARIANTS matrix (ink color × background × which layers are included).

Reuse the engine's primitives — tokenize_path, load_svg_paths, fit_contour, fit_circle, path_el, circle_el, render_svg, write. Don't reimplement maths.

4. Build, compare visually, and tune

Run python build.py to emit the SVG(s), then python scripts/build_gallery.py (point its CONFIG block at this logo's build.py) to write a self-contained gallery.html. Open it and compare to the original: the gallery's Side-by-side / Overlay / Difference views make mismatches obvious — in Difference mode aligned ink cancels to black and any deviation glows. Its per-contour tuning table maps each visible wobble to the knob to turn. Adjust each contour's (step, passes) and re-run build.py + build_gallery.py:

• too rough / jittery → raise passes or step;
• lost detail / too round → lower step and/or passes.

See references/tuning-guide.md. Iterate until faithful. This is the one step that benefits from the user's eye — show them the gallery.

5. Fan out the variant matrix

Extend VARIANTS for the color schemes and layouts the brand needs (e.g. black/white ink × white/black/transparent background × full-lockup/mark-only). One math definition, many ready files. Optionally add a single theme-adaptive row with ink="currentColor" + background=None — an inline-SVG whose color follows the host's CSS (great with prefers-color-scheme). It's additive (collapses only the ink axis, inline-only), not a replacement for the matrix — see references/layer-structure.md.

6. Rasterize to PNG / ICO / favicon

Point the CONFIG block in scripts/rasterize.js at the logo's SVGs, then npm install (first time) and npm run rasterize. Produces square transparent PNGs, wide PNGs, a multi-size favicon.ico, an Apple touch icon, and web-manifest icons.

6.5. Generate showcase images (optional — needs network + an API key)

For presentation-ready brand renders, composite the mark onto premium backgrounds with scripts/generate_showcase.py (Gemini / Nano Banana). This is the only step with external dependencies: cp .env.example .env and add GEMINI_API_KEY, then pip install -r requirements.txt. Feed a transparent raster variant from step 6 (matching the background's polarity):

python scripts/generate_showcase.py "Brand Name" out/raster/monkey-black-512.png \
    --all-styles --output-dir assets/showcase

The prompt forces the model to reproduce the mark faithfully (recolor + composite only — never redraw). Pick styles from references/showcase-styles.md. Skip this step entirely if you don't want the dependency — the rest of the skill is unaffected.

7. Run the quality checklist, then lay out a clean repo

Before delivering, walk references/quality-checklist.md — the fidelity gate (curve sanity, holes, circles, variant consistency, faithful rasters/showcase). Then organize as assets/ (the SVG + raster + showcase outputs), src/ (build.py + the scripts), archive/ (the original drawing), docs/ (a short math write-up — model it on references/method.md), the gallery.html, and a README.md. Everything regenerates from build.py (+ build_gallery.py / rasterize.js), so the repo stays fully reproducible.

References (read when you reach that step)

• references/method.md — the full mathematical method and why it stays smooth & full.
• references/layer-structure.md — contour roles, holes, the even-odd knockout, circles, currentColor.
• references/tuning-guide.md — choosing (resample_step, smoothing_passes); fidelity vs smoothness.
• references/bitmap-input.md — turning a scan/photo into a polyline SVG first.
• references/quality-checklist.md — the fidelity gate to run before delivery (step 7).
• references/showcase-styles.md — the 12 premium backgrounds for the optional showcase (step 6.5).

Common mistakes

• Replacing the outline with primitives. Don't snap the silhouette to a perfect circle — fit along the drawing; that's what preserves the character.
• Relative/arc path commands. The parser expects absolute M/L/H/V/C/Z. Flatten the SVG first if it has lowercase commands or A arcs.
• Forgetting even-odd. Inner contours only become see-through holes when their group is one path with fill-rule="evenodd".
• Tuning blind. Always compare to the original visually (the gallery) before locking parameters.
• Editing the engine per logo. Per-logo specifics belong in build.py, never in scripts/vectorize.py.
• Letting the showcase model redraw the logo. The showcase step must reproduce the vectorized mark faithfully (recolor + composite only); if a render alters it, regenerate.
• Treating currentColor as a replacement for the variant matrix. It only collapses the ink axis and works inline-only — keep the explicit black/white/transparent files.