Image to design md

Image(s) → DESIGN.md

Take 1 or more raster images (PNG / JPG) and produce a DESIGN.md that conforms to the Google design.md spec. The same workflow used in figma-to-design-md Path C (manual fallback) is the default here, since raster is all we have.

When this skill applies

Trigger on any of:

• The user pastes 1+ image paths and asks for a design system / DESIGN.md.
• The user describes images they have (Pinterest pins, Dribbble shots, competitor screenshots, hand-drawn mockups, scanned printouts).
• An explicit request: "estrai design da queste immagini", "make a DESIGN.md from this image", "build a design system from this Figma export PNG".
• The orchestrator (dev-flow) routes here from phase prd_drafted or empty when the user has no Figma URL but has images.

If the user has BOTH a Figma URL AND images, prefer figma-to-design-md (more precise — Variables panel beats vision LLM) and pass the images to screenshot-to-page later.

What you produce

In dev-flow mode: <root>/.workflow/DESIGN.md + <root>/.workflow/screenshots/<slug>.png for each input image. In standalone mode: <project-root>/DESIGN.md + <project-root>/screenshots/.

The DESIGN.md follows the same 8-section structure as figma-to-design-md produces — same spec — see references/spec.md for the canonical Google design.md spec.

Dev-flow contract

This skill participates in the dev-flow workflow. See references/contracts.md for the canonical schema. Key facts:

• Output goes into <root>/.workflow/:
  • DESIGN.md (the main artifact)
  • screenshots/ (input images copied + slugged)
• State is updated by setting meta.json#phase = "design_extracted" (only if the current phase is earlier in the enum), refreshing updated_at, and appending a history entry.
• Standalone mode (no .workflow/ present) is supported — fall back to writing at the user-specified project root.

Prerequisites

• Vision-capable LLM. This skill leans on the model's vision capability for typography identification, component recognition, and layout inference. Claude Opus/Sonnet/Haiku 4.5+ all qualify; ChatGPT-4o and Gemini also work in their respective harnesses.
• Python 3.10+ with Pillow + numpy + scikit-learn (only for quantize_palette.py).

If neither vision nor Python is available, this skill cannot run — refuse and explain. Don't fabricate token values.

Workflow

Step 1 — Locate the images and the output directory

• The user provides one or more image paths (absolute or relative to cwd). Validate each: exists, is a readable file, has extension .png / .jpg / .jpeg / .webp / .bmp.
• Dev-flow mode: if <cwd>/.workflow/meta.json exists, output goes into <cwd>/.workflow/. Read meta.json#project_name and PROJECT.md (if present) for brand context.
• Standalone mode: identify the project root (current working directory unless user says otherwise). Output goes there as DESIGN.md + screenshots/.
• If a DESIGN.md already exists at the target, ask before overwriting.

Step 2 — Capture brand context

The images alone don't tell you "what this brand is for". Ask the user for, at minimum:

1. Project name (used as name: in the frontmatter and as the slug for screenshot filenames).
2. Audience / use-case (1–2 sentences). Used to write the Overview section.
3. Voice / tone if obvious — playful, editorial, austere, etc.

If PROJECT.md exists in .workflow/, reuse it: it already answers most of these.

Step 3 — Extract the palette via k-means

For each input image:

1. Run python3 scripts/quantize_palette.py <image-path> --k 12 --no-crop to get the top 12 dominant colors.
2. The script returns (hex, percentage) pairs. Don't take all 12 blindly — filter:
   • Drop pure-grays unless they're explicitly a UI surface color.
   • Drop colors that appear only in photographic content (skin tones, sky, foliage). These are content, not brand.
   • Cluster near-duplicates (Δhex < 8 per channel).

When you have multiple images, run scripts/aggregate_palettes.py <img1> <img2> … to merge their palettes into a single deduped list ranked by total weighted frequency. This handles the case "5 screenshots of the same app" where the brand palette repeats across images — the cross-image-recurring colors are the real tokens; one-off colors are noise.

Map the surviving colors to spec-recommended names: primary, secondary, tertiary, neutral, surface, on-surface, error (per the spec's "Recommended Token Names"). When in doubt about which color is primary, ask the user — typically it's the most saturated brand-distinct color, NOT the one that takes the most pixels (white usually wins on pixel count).

Step 4 — Identify typography

This is the highest-confidence-loss step. Vision LLMs can identify a font family within a small set of candidates but cannot read fontFamily metadata that doesn't exist in raster pixels.

For each image:

1. Identify text regions and their visual hierarchy (display headline, body, caption, button label, …).
2. For each tier, estimate:
   • Family: name 1–3 likely candidates. If the font is clearly proprietary (Söhne, Roobert, Wise Sans, Helvetica Now, Cera Pro, GT Walsheim, Brown, Söhne Breit), name it AND a Google Fonts open-source fallback (Inter, Manrope, DM Sans, IBM Plex Sans). The DESIGN.md should emit the fallback as fontFamily and document the proprietary original in the prose.
   • Size: estimate from pixel height of cap-letters relative to known image dimensions. Use designer-friendly numbers (12, 14, 16, 18, 20, 24, 32, 40, 48, 64, 72, 96, 126).
   • Weight: identify from stroke thickness — typical buckets are 400 (regular), 500 (medium), 600 (semibold), 700 (bold), 900 (black).
   • Line-height: estimate from spacing between lines.
   • Letter-spacing: estimate, expressed in em (positive ≈ tracked, negative ≈ tight).

State the confidence level explicitly in the prose: display-lg: Inter 96/0.9/900 (visually identified — original may be Söhne Breit, swap if available).

Step 5 — Identify components

For each image, scan for recurring UI patterns:

• Buttons: shape (pill / rounded / sharp), fill color, text color, padding hint, hover-state if visible.
• Cards: radius (sharp ≈ 4px, soft ≈ 12–16px, generous ≈ 24–40px, mega ≈ 9999px-pill — eyeball it), surface color, border presence.
• Inputs: radius, fill, border treatment, error state if any.
• Navigation: top-bar / sidebar / mixed. Active-item style.
• Badges / chips: shape, color mapping (status-driven? semantic?).

Map each to the components block in the frontmatter. Don't invent variants you don't see — if only the default state is visible, emit only the default. Variants get added later by screenshot-to-page when more images arrive.

Step 6 — Identify layout, shapes, elevation

• Layout: grid model (single-column / 2-col / 12-col / asymmetric mosaic). Container max-width if visible. Spacing rhythm if you can read it (4 / 8 / 16 px clusters).
• Shapes: corner-radius scale (collect distinct radii observed). Map to rounded tokens (sm/md/lg/full).
• Elevation: drop-shadows? ring borders? tonal layers? For each card-like surface, decide.
• Brand voice prose: synthesize the Overview section from the user's audience input + observed visuals. 4-6 sentences.

Step 7 — Synthesize and write the DESIGN.md

Assemble the frontmatter (colors / typography / rounded / spacing / components) + 8 prose sections (Overview / Colors / Typography / Layout / Elevation & Depth / Shapes / Components / Do's and Don'ts) per the spec.

Validate before writing:

• Frontmatter delimited with --- lines.
• All color values are "#rrggbb".
• All dimensions use px, em, or rem.
• All {path.to.token} references resolve.
• 8 sections in spec order, no duplicates.
• colors.primary is defined (mandatory).

Write to <root>/.workflow/DESIGN.md (dev-flow) or <project-root>/DESIGN.md (standalone).

Step 8 — Copy images to screenshots/

Each input image is copied into <root>/.workflow/screenshots/<slug>.png so:

• Future screenshot-to-page invocations can use them as references.
• The DESIGN.md prose can cite them: "Pattern observed in screenshots/dashboard.png".

Slug each filename: lowercase, replace non-alphanumeric with hyphens, ensure uniqueness.

Step 9 — Update state and report

Dev-flow mode: before reporting, update <root>/.workflow/meta.json:

• if current phase is earlier than design_extracted, set phase = "design_extracted".
• bump updated_at to ISO-8601 UTC now.
• append a history entry:

  {
    "skill": "image-to-design-md",
    "ran_at": "<now>",
    "inputs": {"image_count": <n>, "images": ["<paths>"]},
    "outputs": ["DESIGN.md", "screenshots/"],
    "phase_before": "<prev>",
    "phase_after": "design_extracted"
  }
  

Then tell the user, briefly:

• Number of images analyzed.
• How many color / typography / component tokens you extracted.
• Confidence per section (palette: high; typography: medium-low; components: medium).
• Sections you flagged as low-confidence and would benefit from human input.
• Where the file was saved.
• In dev-flow mode: the new phase and the next-step proposal (design-md-to-app to scaffold the codebase).

Important constraints

• Don't invent values you don't have evidence for. If only one image is supplied and it doesn't show typography, emit a stub typography: block with a prominent note in the prose, OR ask the user for one extra image showing text. Don't fabricate.
• Be explicit about font names. Vision can't read embedded font metadata; everything is a guess. Always name a Google Fonts fallback and mention the likely original in prose.
• Don't try to identify dark/light modes from a single image. If the image is dark-themed, generate the dark variant; auto-derive light only if the user requests.
• Don't store images outside screenshots/. Skills downstream expect them there.
• Don't process images larger than ~10 MB without warning the user — k-means on a 4K image is slow.
• Per la lingua delle prose: la skill scrive le sezioni in inglese di default. Se il PROJECT.md o l'input utente è in italiano, la skill rispetta la lingua dell'utente per la prosa (frontmatter resta in formato spec — nomi token in inglese).

Quality bar — what a good output looks like

• Palette section: 8–15 named tokens with prose justification ("primary #7839CD reserved for the load-bearing CTA"). No "color1 / color2" placeholder names.
• Typography section: 6+ levels with samples that read as real product copy (extracted from PRD.md if available, or asked from the user). NEVER use Lorem ipsum or "Body 16/400/regular" as samples.
• Confidence flags: every section that contains a guess is marked. Future agents reading the DESIGN.md should know what to trust.
• Screenshot citations: the prose references screenshots/<file> for any inferred-from-image claim.

A good DESIGN.md from this skill is one that the user can read, agree with, edit a few prose lines, and ship — not one they have to rewrite.