A theme is six tokens, not a redesign

The risk in this brief is named in the brief: personalisation that becomes a collection of unrelated websites. So the answer is a hard boundary rather than taste. A theme may change exactly six declared tokens. Everything else is the constant.

All three theme pages are rendered by one file, theme.js, from the same block sequence in the same order. Only the words and the six tokens differ — which is why the check below is possible at all: the rendered class sequence of each state is asserted identical across all three themes, and so are the lane and composer geometries.

1The contract

2Tokens and accents

Every accent is a pair — the accent carries rules and fills, the ink carries text

All three hues sit in the shipped warm room on #F7F3EB. That is the second guarantee against three-unrelated-websites, and it is why none of them is a blue or a violet — a cool accent on this paper reads as a different product, not a different expert.

3The eight answer block types

Same markup on every theme · what varies is listed per row

4Assignment — a theme is derived, not chosen

This is the part that makes it authentic rather than decorative, and it comes from the 2026-05-28 expert response material audit rather than from taste. That audit found citation behaviour varies enormously across the published roster: source visibility ranged from 3/7 to 6/7 related answers, and library depth from 1 material and 17 chunks at one end to 10 materials and 511 chunks at the other.

• Editorial — a named published work and a habit of citing it. The audit's strongest pattern: the expert whose answers name her own book by title.
• Playbook — material that is frameworks, diagnostics and sequences; the answer is a procedure.
• Coach — reflective, first-person material in wellbeing or personal development.

5Accessibility

• Any theme token used for text is the -ink one, and all three clear 4.5:1 on both #F7F3EB and #FFFFFF. The raw accents are 3.9–4.5:1 and are restricted to rules, fills and markers — the shipped system already works this way, which is why --terra-ink exists alongside --terra.
• Colour never carries meaning alone. The numbered citation marker is a number; the source pill is text; the fallback is a dashed border and a sentence.
• Quote blocks are <figure> + <figcaption>, so the attribution is associated rather than merely adjacent.
• The image placeholder carries role="img" and a real aria-label describing the diagram's content — not "diagram".
• Focus rings use --accent-ink at 2px with 3px offset, on every theme.
• Coach suppresses the inline marker. The citation is still in the DOM in the footer — a screen-reader user loses nothing, and the assertion suite checks the source list is present on all three themes.

6Mobile

• Lane is full-width with 16px gutters; every block reflows, none is hidden.
• Video is the only block that changes layout — 132px side thumb becomes a stacked full-width thumb. Same markup, one grid change, identical on all three themes.
• The booking CTA goes full-width and drops its left auto-margin, so it is never a small target in the corner.
• Rhythm multipliers apply on mobile too — coach stays 14% airier, which is the point of it.
• Quote sizes are unchanged; they are already inside a comfortable measure at 390px.

7Fallback rendering

Two failure modes, and both are real rather than hypothetical — the audit found published experts with as little as one material and thirteen chunks.

• No media. Never render an empty card or a grey box. The answer states there is no material, names what the expert can speak to, and offers that instead. Drawn as state 5 on all three theme pages.
• No citation available. The source footer is omitted rather than shown empty — but the answer must say the claim is unsourced. Silence plus a confident tone is the failure this product cannot afford.
• Thin library generally. An expert below roughly two materials will hit the fallback often, so the fallback must ship with the first theme, not after it.

8Recommendation — prototype Editorial first

• It is the house hue. Terracotta and Fraunces are already shipped, so theme 1 adds zero new brand surface — it is the lowest-risk thing to build and it doubles as the default for any expert without a classification.
• It exercises the citation path hardest — inline markers plus a source footer. The audit's weakest finding was uneven source visibility, so building the most citation-demanding theme first surfaces pipeline gaps immediately, rather than after two themes are in production and the fix is three times the work.
• It covers most of the roster on day one. Six of the ten audited experts already behave editorially, naming titles in their answers.
• The other two become token deltas, not builds. Once the eight blocks exist and the fallback is proven, Playbook and Coach are six values each.

9What this is not

• No code and nothing deployed to the product. These are design mocks on the shipped token set; the shell CSS here is a copy for presentation, not a patch.
• The experts on the theme pages are invented. The published experts in the audit are real people, and putting fabricated quotes or invented book titles in their mouths for a mockup is not a shortcut worth taking. The audit is cited above only for what it observed.
• Nothing outside the answer lane is proposed — navigation, routing, profile and the booking flow itself are untouched.

See theme 1 →