Web art direction
Components decide whether a page works. Composition decides whether anyone remembers it. This skill handles the composition: concept, first screen, section rhythm, type, color, imagery and copy. Then a critic reviews the result until it ships.
Tools
The design-refs MCP server exposes refs_api and refs_run (code mode). Call refs_api once per session to load the API. If those tools are missing, tell the user the design-refs server is not connected, and point them to the plugin README. If DESIGN_REFS_DIR is set, read refs/index.json, lessons.json and the images from that folder instead.
Load these files only when their step comes up:
- references/moves.md: composition moves with tier-A references
- references/page.md: rhythm, type, color, imagery, copy, mobile
- references/slop.md: the generic look, with counter-examples
Process
1. Brief
Find out who the page is for, what they must believe, the one action, and which real assets exist (logo, photos, product, copy, prices). If any of these is missing and can't be inferred, ask once, with at most three questions. Never invent customers, logos, reviews, prices or stats. Use marked placeholders and list them at the end.
2. Memory
Run lessons.list() and read every lesson. They come from past critiques and are binding unless the brief overrides one. Name the lessons that apply to this page.
3. References
In one refs_run call:
- three tier-A references from the page's category,
- two tier-A references picked at random (
Math.random) from other categories. These are challengers against the category default, - two tier-C references from the category, as what to avoid.
Look at their first screens. Measure the most relevant one live with refs.measure(url). For each, write one line on what you take from it: a decision, not a layout. If the user names sites they like, refs.capture() them first and tag them.
4. Direction
Read moves.md. Write down:
- The rut: the page this category always ships, and the anti-slop version of it (e.g. cream background + serif + terracotta). Both are off-limits unless the brief pins them.
- Thesis: one sentence naming the one idea this page owns.
- First screen: one first-screen move and exactly what sits where, at what scale.
- Sections: four to six section moves in order, with the rhythm (dense or quiet, container, scale) of each.
- Type, color, imagery: one type move, one color strategy with hex values, and the imagery plan with real sources.
Put this contract in an HTML comment at the top of the page. If any line reads like a mood instead of a decision, it isn't decided yet.
5. Build
Read page.md and slop.md, then build the contract with full commitment: real fonts, real images (check that each URL loads), real copy. Use the ui-libraries-no-ai-look skill for interactive components. Fix the hero first, then everything below it.
6. Shoot
node <this-skill-dir>/scripts/shoot.mjs <page.html-or-url> <out-dir>
This writes desktop.jpg, desktop slices, mobile.jpg, mobile-full.jpg and report.json with mechanical tells (eyebrows, gradient text, purple gradients, equal card rows, stock phrases, timid h1, broken images, overflow). Fix every tell the brief doesn't justify before the critique.
7. Critique loop
Spawn the critic with a fresh context, never a fork of this conversation:
- if the
web-art-directoragent type is available (it ships with this plugin), use it; - otherwise spawn a general-purpose agent whose prompt starts with the full text of
agents/web-art-director.mdfrom this plugin.
Send it these inputs:
- the brief and the user's answers,
- the contract,
- the page path,
- the shot folder and
report.json, - the reference ids from step 3.
It returns rebuild, fix or ship, with ordered fixes, and records lessons itself through lessons.add.
rebuild: go back to step 4 for the named regions.fix: apply every fix in one batch, re-shoot, and send the new shots to the same critic (continue it, don't spawn a new one) for a verdict.- Repeat until
ship, at most three rounds. If round three still isn'tship, show the user the open items and let them decide.
Report the critic's verdict word exactly as it wrote it, never a softer one.
8. Close
Tell the user:
- the thesis,
- the final verdict,
- the lessons that were added or reinforced,
- which placeholders need real content.
Dashboards and app UI
Usability beats expression here. Use moves only for the shell's identity (type, color, one signature detail). Let ui-libraries-no-ai-look drive the components, and still run steps 2, 6 and 7.