Josef Müller-Brockmann (1914–1996), Zurich; Grid Systems in Graphic Design (1981) is the corpus. The grid is treated as an ethic, not decoration: “The grid system is an aid, not a guarantee. It permits a number of possible uses and each designer can look for a solution appropriate to his personal style. But one must learn how to use the grid; it is an art that requires practice.” This skill encodes that discipline AND — the part most attempts get wrong — the front-end engineering to make the grid genuinely load-bearing on the web, plus a harness that PROVES it.

HARD GATE — Do NOT ship a grid page without running verify_grid.js first. A visually plausible layout is not evidence of grid adherence — the harness is. Zero failing assertions is the only passing bar.

Two real review notes this skill exists to prevent:

1. “the grid is just slapped on top and misaligned” → the overlay wasn’t in the same content box as the content (see §2.2).
2. “the H in the headline is off the grid” → the headline’s BOX was on the grid but its INK wasn’t; large glyphs carry a side-bearing (see §2.6). Box-on-grid ≠ ink-on-grid.

• Objective order. The grid brings “constructive thought,” legibility, and “objective and functional” design. Restraint is the point; the system, not the ego, organizes the page.
• Modular grid. Divide the type area into a field of modules — columns AND rows — separated by consistent gutters, inside defined margins. Text and images occupy whole modules. Müller-Brockmann specimens common field counts (8 / 20 / 32 fields). For the web, a 12-column grid + 8px baseline is a robust general default; a 6×6 or 4×8 modular field grid when you want visible rows too.
• Baseline grid. Vertical rhythm is sacred: leading = a whole multiple of the baseline unit, and every element snaps to it. This is what makes facing columns and images line up across the page.
• Typography. A grotesque sans (Akzidenz-Grotesk / Helvetica; on the web Inter, Helvetica Now, Archivo). Flush-left, ragged-right. Few sizes, large jumps in scale for hierarchy; objective, not expressive. Big numerals/data set large is a signature move.
• Palette. Pure white paper, near-black ink, one accent — red is canonical. Avoid the warm-cream “Claude look”; never blue/purple gradients (hard house rule).
• White space + asymmetry. Generous margins; asymmetric compositions held in tension by the grid.

grid_tokens.py emits this whole scaffold correctly; the rules below are why it’s built the way it is.

Put every grid parameter in :root CSS variables — --cols, --gutter, --margin, --bl (baseline), --lh (leading=3×bl), --maxw. Content and the overlay both read these same variables. Never hand-author the overlay separately or it will drift.

Failure mode: content sits in a centered max-width container while the overlay is a full-width sibling of the section. On any viewport wider than --maxw, the centered content and the full-width overlay no longer share column positions → “slapped on top / misaligned.” Fix: put .guides inside the same .wrap, and draw the column guides with left/right = var(--margin) and the same repeat(var(--cols),1fr) + column-gap:var(--gutter). Then the overlay columns are the content columns at every width. Add left/right margin lines at var(--margin).

Don’t eyeball spans. Each horizontal band spans all columns and re-exposes them:

.band{grid-column:1 / -1; display:grid; grid-template-columns:subgrid; column-gap:var(--gutter); align-items:start;}
@supports not (grid-template-columns:subgrid){ .band{grid-template-columns:repeat(var(--cols),1fr);} }

Children place with grid-column: <startline> / <endline> (e.g. 1 / 6, 6 / 13). Every headline, paragraph, photo, caption now snaps to identical lines.

• Leading = --lh (e.g. 24px = 3×8). Every line-height a multiple of the baseline, in px (not unitless) for display type — unitless line-heights on large type push the box off the grid.
• Every margin/padding a multiple of the baseline. Spread top/bottom padding a multiple too, so content starts on a line.
• Media heights = multiples of the leading (e.g. 240/360/432/480px) so a photo’s top AND bottom both land on lines.
• Hairline rules sit inside a baseline-height band, not free-floating.

A control (button + G key) toggles body.grid-on; overlay fades 0→1. Overlay draws: translucent numbered column fields, the baseline (major line every --lh, faint minor every --bl), and margin lines. Showing the real grid the page is built on IS the demo.

A 180px headline whose layout box is exactly on line 1 still looks misaligned against body text, because the letterform’s ink is inset by its left side-bearing. Cure at runtime:

// after document.fonts.ready and on resize:
var cvs=document.createElement('canvas'),ctx=cvs.getContext('2d');
document.querySelectorAll('.masthead,.numeral,.shead h2,.h2b').forEach(function(el){
  el.style.marginLeft='0px';
  var cs=getComputedStyle(el),ch=(el.textContent||'').trim()[0]; if(!ch) return;
  if(cs.textTransform==='uppercase') ch=ch.toUpperCase();
  ctx.font=cs.fontStyle+' '+cs.fontWeight+' '+cs.fontSize+' '+cs.fontFamily; ctx.textAlign='left';
  var abl=ctx.measureText(ch).actualBoundingBoxLeft;     // +ve = ink overhangs left of box
  if(isFinite(abl)) el.style.marginLeft=abl.toFixed(2)+'px'; // shift box so INK lands on the line
});

Apply to the masthead, big numerals, and section headlines. It scales with fluid type (re-runs on resize) and uses the actually-loaded font, so it’s correct in the user’s browser. CRITICAL measurement caveat: side-bearing is font-specific. If you measure with the wrong font you get the wrong nudge. Headless/sandbox Chrome usually lacks the webfont, so canvas falls back to a different grotesque (measured −16px on the fallback vs −7px on real Inter for the same H). To verify optics offline you must embed the real webfont via @font-face (local TTF). In production the runtime JS measures the loaded font and is correct.

Render with headless Chrome (Puppeteer) and assert, at several widths including > and < --maxw (to catch centered-container drift, e.g. 1440 / 1180 / 900):

1. Column adherence — every placed .band > * left snaps to a column START and right to a column END (~0px). Exclude the optically-aligned display elements from this box check (their box is intentionally side-bearing-offset; they’re validated in step 4). Gotcha: build BOTH the column-start set and the column-end set — a grid item spanning “to line N” ends at the far side of the gutter, so single-edge math falsely reports a one-gutter error.
2. Overlay match — each .guides .col rect equals the computed column rect (~0px).
3. Baseline — text tops modulo the baseline ≈ 0 (tolerance ≈ half a baseline; the box-top is a proxy — the leading does the real work).
4. Optical ink — each display element’s ink-left (box − actualBoundingBoxLeft, real font) equals its own column line (nearest column-start to its box), not always line 1.

Sandbox Chrome flags that work: --headless=new --no-sandbox --disable-gpu --disable-dbus --use-gl=angle --use-angle=swiftshader. file:// works for non-ES-module pages; the CLI --screenshot can hang on tall pages — drive via Puppeteer and screenshot per viewport. Read PNGs back with the image-capable Read tool to eyeball a zoom crop of the top-left corner (masthead vs body vs column line) — the fastest human check.

A clean run looks like: col=0px overlay=0px baseline≤4px ink=0px → GRID VERIFY: PASS.

• Palette: white #fff, ink #111, one accent (Swiss red #e4002b). No warm-cream Claude look; no blue/purple gradients.
• Type: a real grotesque webfont (Inter / Helvetica Now / Archivo) for display + body; a mono (Space Mono / IBM Plex Mono) for folios, captions, grid annotations — reinforces the technical register. Non-Latin via Noto Sans JP etc.
• Hierarchy through scale + weight + white space, not color. Treat key data as large numerals. Kicker labels in mono caps. Per-spread folios.
• Real photography. Ground real subjects in real photos (SearchImages). Host each image via PublishFilePublicly and embed the pub.hyperagent.com URL — a PublishWebpage artifact runs in a sandboxed iframe that can’t authenticate thread-scoped /api/files/... URLs (broken-image trap).
• Type fidelity if you ever rasterize art (cairosvg / headless screenshots / image-gen reference): a Helvetica/Arial CSS stack silently falls back to Noto Sans (reads like Calibri). Render in Liberation Sans or an embedded Helvetica/Arimo TTF before trusting it. (Same trap as the optical-measurement caveat: wrong font in → wrong result out.)
• Spread model: full-width sections, each its own per-spread .grid + .guides, consistent margins/folios.

1. Pick the subject; gather real photos; host them publicly.
2. Generate the scaffold: python3 grid_tokens.py (or --scaffold for a full page; --cols/--baseline/--gutter/--margin/--maxw/--accent to taste; it warns if gutter/margin aren’t baseline multiples).
3. Build spreads as subgrid bands; place everything by column line; lock spacing/line-heights/media heights to the baseline.
4. Add the overlay (same content box) + toggle + optical-alignment JS (already in the scaffold; point its selector list at your display elements).
5. Publish, then verify: CHROME=… PUP=… node verify_grid.js <file-or-url> --widths=1440,1180,900. Eyeball a top-left zoom crop. Fix, republish.

• grid_tokens.py — deterministic scaffold generator. Emits the :root tokens, .grid/.band (subgrid) scaffold, .guides overlay CSS, toggle JS, and the optical-alignment JS — all wired to one source of truth. --scaffold emits a full minimal HTML page. No network/credentials.
• verify_grid.js — Puppeteer harness implementing all four checks above with the corrected both-edges column math, the optical-exclusion, per-element column-line ink targeting, and PASS/FAIL output at multiple widths. Env: CHROME (chrome binary), PUP (puppeteer-core module path).

A grid you can’t toggle on and measure is a mood board, not a system. Build it from one source of truth, prove it at 0px, and align the ink.