New UI mock

For work with no reference screenshot to copy. If there is one, use clone-prototype instead. Measurement beats invention every time.

Everything renders on the canvas from mockups/canvases/<slug>/; see prototype-canvas for running it, and its references/layout.md for the folder and layout.json rules.

The plugin ships the template folder and every worked example. sp-canvas and refkit are separate — the plugin cannot run an installer of its own, so if they are not on PATH: uv tool install "git+https://github.com/ReScienceLab/super-prototyping#subdirectory=tools". sp-canvas root then prints where the plugin landed:

KIT="$(sp-canvas root)"
mkdir -p mockups/canvases                       # first board in a project
B=mockups/canvases/<slug>
cp -r "$KIT/mockups/canvases/templates" "$B"
python3 "$B/gen.py"

Everything a run derives goes under $B/scratch/: shots, montages, candidate boards. A -o without a directory writes into whatever the current directory happens to be, which is the user's project root as often as not. The template's parked reference is ref-01-screen.html, and that stays out of git too. Once per project:

for p in 'ref-*.html' '**/assets/refs/' 'scratch/'; do
  grep -qxF "$p" .gitignore 2>/dev/null || echo "$p" >> .gitignore
done

---

1. Find the tokens before you design

Never invent a palette when the product already has one.

• Extending an existing board? Reuse its 00-design-tokens.html block verbatim. Copy it byte-identically into the new file; a sandboxed iframe has no shared stylesheet.
• Cloning a real app's look? Stop and run clone-prototype Phase 1 and Phase 2 first; come back with a measured token block.
• Genuinely new product, nothing to measure? Copy the shipped template folder, change NAME and the prefix, and pick deliberately: a platform-native stack, a neutral ramp, one accent, one danger. Keep the evidence table and write why in it ("iOS system blue", "brand hex from the logo"). An unexplained hex is a future bug.

A new token is a decision, not a convenience. If a screen needs a colour or a size that is not in the block, either it belongs in the block (add it there, in every file) or the screen is wrong.

---

2. Ground the content

Readers take the copy, numbers and states in a mockup as product decisions.

• Take strings from the real source when it exists: localization files, existing screens, the spec. Never from imagination.
• Use real assets over hand-drawn approximations: the actual icon, the actual logo, the actual empty-state illustration, embedded as a data: URI.
• Design the unhappy states too: empty, loading, error, long string, longest plausible number. A mock that only shows the happy path hides exactly the layout problems worth finding now.

---

3. One generator, one row of screens

Same rule as cloning. Write one script that emits every .html in the folder, and edit the script, never the output.

TOKENS = ":root{...}"                      # one source of truth
def page(title, extra_css, body): ...      # TOKENS + frame + body

Constraints (from the canvas renderer):

• Fully self-contained. sandbox="" means no external CSS, JS, fonts or images; data: URIs and inline SVG only.
• Artboard box is 478 × 980. Overflow clips silently.
• iPhone frame 393 × 852 pt at 1pt = 1px: 54px status bar, 125 × 36 Dynamic Island, 139 × 5 home indicator.
• Avoid SF Symbols private-use glyphs; they render as tofu without SF Pro.
• Accessibility is not a mockup detail to skip: real contrast on real backgrounds, ≥ 44pt tap targets, aria-label on icon-only controls. A mock that fails contrast ships a screen that fails contrast.

Proposals to compare

When the ask is "show me some options", make each proposal a whole artboard in one row, not a fragment:

{ "title": "Card layout proposals", "numbered": true,
  "files": [
    { "file": "10-cards-continuity", "label": "Continuity" },
    { "file": "11-cards-structured", "label": "Structured" },
    { "file": "12-cards-reference",  "label": "Reference" }
  ] }

Three is usually the right number. Two reads as a false binary, five as indecision. Make them genuinely different approaches, not three spacing values, and put a one-line rationale at the bottom of each board. Rows align column-for-column, so a second row of the same three under a different state reads as a matrix.

---

4. Iterate against annotated screenshots

For review, the user marks up a screenshot of the canvas with boxes, arrows and numbers, then pastes it back.

1. Echo what you read each annotation as, before touching anything.
2. Change the generator, not the artboard.
3. Re-run the generator; the canvas reloads onto the new board.
4. Verify that region visually before claiming it is done.

Answer every annotation, including the ones you disagree with. Say so in a line and make the change, or say why you did not and what you did instead.

---

5. Verify by rendering

refkit shoot "$B"/*.html -o "$B/scratch/shots" --scale 2
refkit montage "$B"/scratch/shots/*.png -o "$B/scratch/board.png" --height 520

Read the montage. Check, in order: nothing clipped; text wraps where you intended; the same element has the same inset on every screen; nothing uses a colour that is not in the token block (grep -o '#[0-9A-Fa-f]\{6\}' on the generated files and compare the set against :root).

Re-render after every pass. A change you have not re-rendered is not done.