Layout HTML

One text in, one .html file out. The page must survive three tests:

1. It opens from file://, network unplugged, and looks finished.
2. Every figure traces back to a sentence of the source text.
3. A reader who never sees the original loses no claim, no caveat, no number.

Time budget: one reading pass, one question, one file. No research detours.

Hard rules

• No new research. The source text is the entire knowledge base. Never WebFetch, never fill a gap from memory. Missing context stays missing (see Failure modes).
• Figures are quotations. Every SVG visualizes data that exists in the source. A number absent from the text never appears in a figure. No numbers in the text → structure figures only (flow, timeline, hierarchy) or none. Never a decorative fake chart.
• Self-contained or nothing. One file: inline CSS, inline SVG, system font stack. No CDN, no webfont, no <script src>, no external image. The only outbound links are sources the text already cites.
• Fidelity over beauty. Never alter a claim, drop a caveat, soften a hedge, or reorder an argument. Abridge only with a visible [abridged — full text in source] marker.
• Zero JavaScript by default. CSS does the layout. Single exception: Magazine mode may embed ≤30 lines of vanilla JS for a collapsible TOC.
• AskUserQuestion: one question per call; title Layout — {label}.

Step 0: Locate the text

First match wins:

• Study has several layoutable texts (notes.md, council.md, research.md) → one AskUserQuestion (Layout — Source) listing them; never guess.
• Ad-hoc slug: derive from title, lowercase-hyphens (AGENTS.md slug rules); if output/layout/<slug>/ exists, ask before overwriting.
• Source under ~300 words → warn: typography-only card, no figures (a figure needs material).

Step 1: Figure ledger

One reading pass. While reading, hunt figure candidates:

For each candidate, record in a working ledger (in conversation, not a file): the exact source sentence, its section, the figure type. Then cut hard: keep 3–7, ranked by "would the reader remember this tomorrow?". Fewer than 2 solid candidates → the page is typography-only; say so rather than forcing weak figures.

The ledger is the traceability contract: at delivery, every figure cites its ledger line.

Step 2: Art direction — one question

AskUserQuestion — title Layout — Direction:

• Éditorial (recommended) — calm single column (68ch), strong typographic hierarchy, figures interleaved at their reference point. For reading and rereading.
• Magazine — full-bleed header, standfirst, pull quotes, denser figure rhythm, collapsible TOC (the one JS exception). For sharing and skimming.
• other

Never ask about colors, fonts, or spacing — that is the skill's job, not the user's.

Step 3: Design tokens

The <style> block starts with tokens; every later rule uses tokens only:

:root {
  --paper: …;  --ink: …;          /* base pair, contrast ≥ 7:1 */
  --accent: …;                    /* ONE hue, chosen from content mood */
  --muted: …;                     /* captions, meta — ≥ 4.5:1 on paper */
  --step--1 … --step-3: …;        /* type scale, ratio 1.25 */
  --space: 8px;                   /* all gaps are multiples */
  --measure: 68ch;
}
@media (prefers-color-scheme: dark) { /* swap paper/ink, re-check accent */ }

• One accent hue. Two is a design failure.
• Fonts: system stack only (ui-serif headings / ui-sans-serif body, or the inverse per mood). No downloads.
• Accent choice comes from the content (biology → green family, hardware → steel blue…) — decide once, state it in the delivery note.

Step 4: Page skeleton

Semantic HTML5, in order:

1. <header> — <h1> (source title), standfirst (the source's own one-line thesis — quoted, not invented), meta line (source path, date, "layout only — no new research").
2. <nav> TOC — only if the source has >4 ## sections; anchor links; collapsible in Magazine mode.
3. <main> — sections mirror the source structure 1:1 (same order, same hierarchy). Body text in the --measure column, line-height: 1.6.
4. Each figure: <figure> + <figcaption>, placed at the point where the source discusses it — never grouped in an annex.
5. <footer> — provenance: source file, generation date, /layout-html mention, and the source's own bibliography/links if present.

Step 5: SVG recipes

Every figure obeys:

• viewBox set, width: 100%, no fixed pixel height in the HTML.
• role="img" + <title> + <desc> (the desc is the ledger sentence).
• Text inside the SVG ≥ 12px at base scale; colors from tokens only (currentColor, var(--accent)).
• <figcaption> = one factual line + (§ section name) for traceability.
• Geometry hand-coded — no chart library, no generator.

Step 6: Self-check (before Write)

Run against the finished HTML string — fix, never ship warnings:

• <script count → 0 (Éditorial) or exactly 1 inline TOC block (Magazine)
• http matches only inside the sources/footer section
• every <svg has viewBox, <title>, <desc>
• every number appearing in a figure re-greps in the source text
• every source caveat/hedge sentence is present in the page
• [abridged] markers everywhere content was cut
• single <h1>; heading levels never skip
• file ≤ 200 KB

Step 7: Deliver

1. Write the file to the Step 0 contract path.
2. In chat, ≤6 lines, no theater:
   • the path + "double-clique le fichier pour l'ouvrir"
   • figures shipped: N figures — bar pair (§2), timeline (§3), …
   • what was abridged, or "rien coupé"
   • accent hue chosen and why (3 words)

Failure modes

What this skill never does

• Research, fact-fill, or "improve" claims
• Decorative charts with invented data
• Multi-file output, external assets, JS frameworks
• Summary theater in the delivery message