Render Check

A surface is done when it looks right where its reader will look at it, not when its source reads right. Source inspection, a green build, a byte-size match, and an HTTP 200 prove that something was served. None of them shows what was rendered.

Core rule

Open the live surface, capture what it shows across the viewports, themes, and states its reader can meet, and report only what the captures show. Say plainly when a statement is source-only inference. Terminal output the reader cannot see is not evidence either: paste the receipts into the reply.

1. Identify the surface and how to open it live

• Web page or app: the dev server, deployed preview, or production URL. If the browser tooling refuses file:// navigation, serve the directory over localhost; do not route around the policy.
• README, docs page, pull-request or issue body: the hosting site's rendered view after the push or post. A local Markdown preview is a different renderer and counts as source-only for the hosted surface.
• Chart, image, PDF, or email: the exported artifact opened in the viewer the reader will use, at the size it will be shown.

2. Declare the coverage matrix

Write down, before capturing, the cells the pass will cover:

The matrix is the pass's own declaration of coverage. The checker enumerates every declared cell that has no capture, so declare the cells the pass will visit and then visit them.

3. Capture every cell

For each cell:

1. Set the viewport width, emulate the theme, put the page in the state.
2. Wait for load. Reload once if sizing looks stale; some apps measure the viewport only at mount.
3. Take a screenshot and save it under a name that carries the cell, with an extension that matches the bytes: PNG bytes as .png, never JPEG bytes named .png or image bytes with no extension. One capture per cell; a retake replaces the earlier file rather than sitting beside it.
4. Record layout metrics from the same render:

({
  viewport_width: window.innerWidth,
  viewport_height: window.innerHeight,
  scroll_width: document.documentElement.scrollWidth,
  scroll_height: document.documentElement.scrollHeight,
  dark: matchMedia("(prefers-color-scheme: dark)").matches,
})

scroll_width greater than viewport_width is horizontal overflow: content the reader must scroll sideways to reach, or that is clipped. Record it; the checker surfaces it.

With Playwright, one cell is:

page.set_viewport_size({"width": 320, "height": 800})
page.emulate_media(color_scheme="dark")
page.goto(url, wait_until="load")
page.screenshot(path="captures/mobile-dark.png", full_page=True)
metrics = page.evaluate(
    "({viewport_width: window.innerWidth,"
    " scroll_width: document.documentElement.scrollWidth})"
)

Any tooling works as long as each capture has a real screenshot file and metrics from the same render. Image bytes returned by a browser extension or an in-app screenshot must be written to disk before they count as evidence.

4. Probe the usual defects

Look for these in the captures, not in the source:

• Overflow at the narrowest width: a horizontal scrollbar, a clipped table or code block, controls pushed outside the viewport.
• Image sizing: a 2x export displayed at its intrinsic size dominates the column; set the display width to the design's 1x size.
• Theme variants: contrast, badges, diagrams, and screenshots that only work in one theme; a host that picks images per theme needs both variants.
• Clipping and overlap: text, buttons, chips, and menus that clip or overlap; overlays that do not fit the viewport; nested scroll traps.
• Stale sizing: a layout measured once at startup that ignores a resize.
• Document structure: tables that must scroll inside their own container, collapsibles that open, embeds that load, anchors that resolve.

5. Write the report and run the checker

Describe the pass in a JSON report. Screenshot paths resolve relative to the report file.

{
  "surface": "https://example.test/preview/",
  "matrix": {
    "viewports": [{"name": "desktop", "width": 1280}, {"name": "mobile", "width": 320}],
    "themes": ["light", "dark"],
    "states": ["default"]
  },
  "captures": [
    {
      "id": "mobile-dark",
      "viewport": "mobile",
      "theme": "dark",
      "state": "default",
      "screenshot": "captures/mobile-dark.png",
      "metrics": {"viewport_width": 320, "scroll_width": 320}
    }
  ],
  "findings": [
    {"id": "F-001", "summary": "Nav overlaps the title below 360px", "evidence": "rendered", "capture": "mobile-dark"},
    {"id": "F-002", "summary": "Print stylesheet hides the sidebar", "evidence": "source-only"}
  ]
}

Every finding declares its evidence. rendered must cite a capture whose screenshot is real image bytes; source-only is allowed and is labeled as inference. From this skill directory, run:

python scripts/render_check.py path/to/report.json

One line per cell and per finding:

• COVERED: the cell has a capture with real screenshot bytes; the line carries its metrics.
• OVERFLOW: that capture's scroll_width exceeds its viewport_width; the line carries the overshoot in pixels.
• MISSING: the cell has no capture, or its screenshot is absent, empty, not image bytes, or named with an extension that disagrees with its bytes (including no extension), or its recorded viewport_width disagrees with the width the matrix declares for that viewport.
• RENDERED: a finding backed by a covered capture.
• REJECTED: a finding presented as rendered evidence without a usable capture behind it, or with no evidence label.
• INFERENCE: a finding declared source-only.
• SUMMARY: the counts and the exit code.

Exit codes:

• 0: every cell is covered, every finding is backed or labeled inference, and no capture measured overflow. The pass is complete and the surface is clean.
• 1: at least one MISSING, REJECTED, or OVERFLOW line. Do not report the surface as verified; complete the pass or fix the surface, then rerun.
• 2: the report is invalid: no matrix, a capture citing an undeclared cell, a second capture for one cell, a repeated id, a malformed metric, or a name, id, or path carrying a control character or a lone surrogate. Fix the report and rerun.

A capture that cites a viewport, theme, or state outside the matrix is an input error rather than bonus coverage. The matrix is the declaration, the captures are the evidence, and the checker reports the difference.

Every receipt is exactly one physical line. Viewport, theme, and state names, capture and finding ids, and screenshot paths print unquoted, so the checker rejects any that carry a control character, a line or paragraph separator, a lone surrogate, or the · field separator; summaries and reasons are JSON-quoted with any such character escaped. No field a report supplies can add, split, or forge a receipt line.

6. Report to the person

Lead with whether the pass was live-rendered or source-only, then the findings, then the receipts. Paste the checker's lines; do not point at a terminal or a folder.

**Rendered check: live, 4 of 4 cells captured, 1 overflow.**

- **Surface:** https://example.test/preview/ at commit `<sha>`
- **Matrix:** desktop 1280 and mobile 320; light and dark; default
- **Browser:** Chromium via Playwright, `prefers-color-scheme` emulated

Findings, most important first:

1. **Mobile overflow (rendered, `mobile-light`)** — the comparison table is
   348px wide in a 320px viewport; `captures/mobile-light.png`.
2. **Hero image at 2x (rendered, `desktop-light`)** — displayed at 1600px in
   an 800px column; `captures/desktop-light.png`.
3. **Dark badge contrast (source-only)** — the badge color is hard-coded in
   the stylesheet; not confirmed in a capture.

Receipts:

COVERED · cell=desktop/light/default · capture=desktop-light · ...
OVERFLOW · cell=mobile/light/default · capture=mobile-light · viewport_width=320 · scroll_width=348 · overshoot=28
SUMMARY · cells=4 · covered=4 · missing=0 · overflow=1 · rendered=2 · rejected=0 · inference=1 · exit=1

Keep the qualifiers that make the pass checkable: the surface and its revision, the widths, the theme mechanism, the browser. Name what was not covered rather than implying completeness. Report what the captures show; whether a finding matters is the reader's call unless they asked for a verdict.

Limits

The checker validates a report about a pass; it does not perform the pass. It cannot open a browser, judge design quality, confirm that a screenshot shows the cell its capture claims, or detect defects that leave no trace in the recorded metrics. Horizontal overflow is the one defect it measures; the rest of the probe list is the reviewer's eye on the captures.