Generated from cat-harness/skills/ui/theming/site-presentation-assets.md — do not edit here.

✎ Edit this page’s source

Site presentation assets — the node is the source, the stylesheet is a rendering

A visual fact has one home, and it is a node in the knowledge graph. A colour, a glyph, an avatar, a theme’s token set: declared once as data, and the stylesheet the site serves is generated from it.

The owner’s requirement, and the whole of it:

named css assets in KG rather than hardcoded colors

Authoring the stylesheet by hand beside the nodes gives two answers to one question, free to disagree — and they do disagree, silently, because nothing compares a hex triple in a .css file with a hex triple in a .ts file. This is the same failure as a hand-written title: in a Jekyll config next to a declared instance name, which diverged the day it was written.

The test

Could a reader change this visual fact in one place and be sure the site agrees?

If the answer needs them to remember a second file, the fact has two homes and one of them is wrong.

Four rules

1 — Never hand-edit a generated stylesheet. It is build output that happens to be committed, and it is committed only so it is readable on the forge. An edit survives exactly until the next generator run.

2 — Gate staleness, do not trust it. The generator takes a --check arm that compares against the committed copy and fails rather than writing. A generator whose --check is not in CI is a generator whose output drifts, and the drift is invisible because the file looks authored.

3 — A published generated asset needs a Tool that maintains it. Otherwise the provenance is unreachable: nothing can answer “which node produced this stylesheet”, and the asset becomes the third case in covered-is-not-reachable — a mechanism nobody can find. Declare the artefact by its published path, and check that something actually publishes it before claiming so.

4 — An asset reference resolves; it is never composed. A page that names a generated asset must reach it relative to its own location, so the same bytes work at the canonical base and under a staging prefix with no configuration. A page that composes the URL from a base needs a different build per environment, and a staging build that differs from what ships is testing something else.

An address may be relativized. A name may not — a JSON-LD @id or a schema $id is an identity, and relativizing it changes what the document claims to be, resolving against whatever URL happened to serve it.

Themes are token sets, not stylesheets with branches

A theme declares tokens — the named values a stylesheet reads. Light and dark are two valuations of one token set, not two hand-maintained blocks that drift apart.

So a new colour is a token on the node, and the generator decides how it reaches CSS. Adding a @media (prefers-color-scheme: dark) branch by hand to cover a value the node does not carry is rule 1 broken with extra steps: the node still does not know about the colour, so the next instance to render it gets the light value in the dark.

The third state applies here too

An asset whose presence cannot be determined — the site not built, a fetch refused — is not present. Report it as could-not-determine and do not let it clear the assets the check did not reach. A generated asset that silently went missing renders as an unstyled page, which looks like a design choice rather than a failure.

What this skill does not cover

• Prose voice and tone — one-voice-style-guide.
• Whether a rendered block LOOKS right — that is a content question, and the rendering auditor’s.
• How a link is spelled — readme-sections §”On link style — raw is not the private-repo answer” (summarised in the agent-memory entry link-style-raw-is-not-the-private-repo-answer).

This skill is only about the direction of authority: the graph decides, the stylesheet reports.

Why this skill exists at all, recorded because the gap was the finding

It was authored 2026-09-20 after two generators — both committed, both published, both single-file — could not be given Tool nodes, because satisfies requires a skill and no skill stated this capability. The mechanism had existed for some time; the vocabulary had not.

That is the third mismatch in covered-is-not-reachable, and it is invisible to bun run cat tools:coverage by construction: that tool enumerates skills and asks which lack Tools, so a capability nobody stated is absent from the list it walks. Bean yean.

The skill was written before the nodes, and deliberately names no script. A skill written to give a command somewhere to point is a Tool with front matter: it passes every check and teaches nothing.

Processes that run this skill

process	step(s) that name it
Code node review	Review the schema definition node