Image to SVG

Reconstruct the supplied image as real vector geometry, using semantic parts and rendered feedback rather than wrapping a bitmap in an SVG. Fidelity to the source is the default; redesign or stylization requires the user's direction. No image-generation model, external API, GPU, or tracing package is required. This skill includes its own scripts and references; no other skill is required.

For standalone conversion, do not invoke a planning or image-generation skill: the supplied image establishes the visual direction. Ask only about consequential source or fidelity ambiguities. When a production workflow supplies a selected image, optional baseline SVG, and a bounded conversion scope, read vector-handoff.md. Return a separate candidate; do not take over art direction or start another production workflow.

1. Inspect the actual source

Open the supplied image with an available image-viewing tool before drawing. Inspect its pixel dimensions, aspect ratio, orientation, transparency, and visible content. Do not infer shapes from a filename or claim to inspect an inaccessible attachment. If the source is missing or cannot be viewed, ask for an accessible image instead of inventing a reconstruction.

Identify the intended artwork separately from any screenshot frame, checkerboard, or surrounding UI. Preserve the source canvas and background by default; do not automatically crop whitespace or turn a white background transparent. Ask only when ambiguity about the target region or fidelity would materially change the result. Preserve the original and write conversion outputs separately.

2. Choose the conversion approach

• Flat logo, icon, or illustration: rebuild meaningful shapes with primitives and short paths. Prioritize silhouette, cutouts, spacing, and layer order.
• Line drawing or sketch: preserve stroke rhythm, junctions, open ends, and intentional irregularity; do not turn every stroke into a filled blob.
• Dense or photographic image: explain the tradeoff between a simplified editable illustration and a high-detail trace. Do not silently substitute a stylized result when close reproduction was requested.
• Explicit automatic tracing: use a suitable available local tracer, inspect its result, and disclose whether the output is contour-editable or semantically organized. Do not claim to have run a tracer when drawing paths manually.

Read conversion-modes.md for tracing, lettering, photographic sources, and low-resolution edge treatment when relevant. Do not install heavyweight tracing stacks or upload the source to a conversion service by default. A task to convert an image does not require generating a new bitmap. If the input is already SVG, inspect and edit that source instead of rasterizing and retracing it. Preserve existing IDs and unaffected parts.

3. Map the source to editable parts

Use a coordinate system matching the source aspect ratio, commonly viewBox="0 0 WIDTH HEIGHT" with the source pixel dimensions. Keep margins and landmark positions proportional; never stretch the image to a square to simplify authoring. Record the target dimensions or crop if the user requested a change.

Identify a small set of recognition-bearing parts, their approximate bounds, colors, holes, attachment points, and back-to-front order. Separate intentional detail from compression artifacts or antialiasing. For multipart work, assign stable IDs such as bird.body, bird.wing, and bird.eye and group related parts. A separate scene file is useful for complex iterative work, not required for a simple conversion. When using one, follow the bundled scene-model.md schema so its mappings can be validated. For a new scene, record observed art direction and semantic parts as a v2 plan, then add actual tags/parents and set phase: "bound" after authoring. Reuse supplied plans and existing IDs; preserve v1 maps without requiring migration. A generated reference must not override explicit preservation requirements from the handoff.

4. Author genuine vector geometry

Create a standalone SVG with the SVG namespace and a finite positive viewBox. Use circles, ellipses, rectangles, and polygons where they fit the reference; otherwise use purposeful Bézier paths. Place anchors at corners, extrema, and inflections rather than following every source pixel. Preserve intentional asymmetry, negative space, overlapping objects, and holes in compound paths.

Keep meaningful parts independently editable, fills and strokes explicit, and resource references local. Use presentation attributes for compatibility with the bundled validator. Gradients are appropriate when visible in the source; they are not a reason to flatten the artwork. Include a useful title and, when needed, desc without inventing facts about the image.

Do not use embedded or linked raster images, data-URI bitmaps, foreignObject, scripts, external resources, or one rectangle per pixel to simulate vectorization. Do not change branding, lettering, color count, or background to make conversion easier. Disclose approximations and font dependencies rather than promising exact reproduction from insufficient source detail.

5. Validate and compare against the source

Check XML, namespace, viewBox, geometry, duplicate IDs, and local references; also check that the output contains no raster content. Resolve SKILL_DIR to the directory containing this file, not the current working directory. Read tooling.md for the bundled helpers, runtime dependencies, static-vector profile, and exit codes. Python 3.10+ is required.

python3 "$SKILL_DIR/scripts/validate_svg.py" converted.svg
python3 "$SKILL_DIR/scripts/render_svg.py" converted.svg --output preview.png --width 768

Use validate_scene.py scene.json for a plan-only contract check. Add --scene scene.json to SVG validation only after a scene is bound; the command rejects an unbound plan rather than pretending its mapping was verified. Choose the preview width for the source or intended display size. Fix validation errors before rendering, then actually open the PNG alongside the source. Do not strip unsupported source features just to satisfy the bundled static-vector profile.

Compare at the same aspect ratio, crop, and background, both at source/display size and zoomed in. Inspect silhouette and landmark alignment first, then holes, occlusion, spacing, palette, stroke weight, and lettering. Check transparency on both light and dark backgrounds. A successful parse or render does not establish visual fidelity, and pixel-error metrics are not a substitute for inspection.

Correct the largest visible discrepancy with a focused edit, revalidate, render, and inspect again. Preserve the best version. Stop when the brief is satisfied or after three unproductive correction rounds; report remaining limitations rather than repeatedly adding detail. Rerender after cleanup or path simplification. The bundled measure_svg.py reports structural complexity; normalize_svg.py only normalizes XML serialization, not geometry. Use compare_render.py for optional pixel measurements, not as a fidelity or design-quality verdict. If rendering or image viewing is unavailable, explicitly state which checks were not performed and do not describe the conversion as visually verified.

Deliver

Return the editable .svg and a rendered preview when available. Briefly state the conversion method, any intentional simplification, text/font treatment when relevant, and unresolved fidelity or verification limits. Keep source images and internal iterations out of the deliverable unless requested.