agentleFS
LengrosLengros/bennu/.claude/skills/design/SKILL.md

design

Skill2 starsChanged 4 months agoMITView on GitHub

Design a distinctive, heart-winning visual artifact — landing/marketing page, one-pager, hero, product page, pitch deck slide, theme — by first committing to a NAMED art direction, then hand-coding self-contained HTML/CSS, rendering it to a PNG via headless Chrome, and LOOKING at it to critique and fix. Use when the user wants something beautiful, on-brand, or visually distinctive — not the safe default look. Its whole reason to exist is to escape the model's beautiful-but-bland default attractor (warm cream + sage/clay + Fraunces serif + soft rounded cards + radial orbs — the "looks like Claude/Anthropic" look). NOT for structural diagrams (use /diagram) or pure copy/IA (use the doc flow). Russian triggers — "сделай красиво", "нужен крутой дизайн", "не дефолтный вид", "запили лендинг", "сделай one-pager красивым", "подбери стиль/палитру", "сделай необычный дизайн", "оформи визуально".

198 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use banner-design, ui-styling and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • seikaikyoSkill

    banner-design

    Design banners for social media, ads, website heroes, creative assets, and print. Multiple art direction options with optional generated or supplied visuals. Actions: design, create, generate banner. Platforms: Facebook, Twitter/X, LinkedIn, YouTube, Instagram, Google Display, website hero, print. Styles: minimalist, gradient, bold typography, photo-based, illustrated, geometric, retro, glassmorphism, 3D, neon, duotone, editorial, collage.

    Also in 38 projects that use this

  • MicrosoftSkill

    ui-styling

    Guide for the visual style, structure, and shared building blocks used by Semantic Link Labs interactive UI tools (HTML widgets and anywidget-based widgets). Use this when adding a new interactive UI, modifying an existing one, or adding shared visual components.

    Also in 39 projects that use this

  • seikaikyoSkill

    brand

    Brand voice, visual identity, messaging frameworks, asset management, brand consistency. Activate for branded content, tone of voice, marketing assets, brand compliance, style guides.

    Also in 43 projects that use this

  • mayankjain1815-commitsSkill

    infra-audit

    Reads docker-compose, env files, ORM configs, and connection strings to map current infrastructure. Flags missing layers (cache, queue, analytics) based on observed access patterns. Outputs a structured infrastructure manifest.

    Also in 3 projects that use this

  • JetBrainsSkill

    slides

    Create and edit presentation slide decks (`.pptx`) with PptxGenJS and render/validation utilities. Use when tasks involve building a new PowerPoint deck, recreating slides from screenshots/PDFs/reference decks, modifying slide content while preserving editable output, adding charts/diagrams/visuals, or diagnosing layout issues such as overflow, overlaps, and font substitution.

    Also in 42 projects that use this

  • AnthropicSkill

    design-system

    Audit, document, or extend your design system. Use when checking for naming inconsistencies or hardcoded values across components, writing documentation for a component's variants, states, and accessibility notes, or designing a new pattern that fits the existing system.

    Also in 49 projects that use this

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 174 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When you catch your draft using ≥2 of these rows *and nobody asked for that look*, you are in the attractor.

    line 37
  2. Distinctive:

    Before any drafting, propose 2–4 named, genuinely distinct directions.

    line 72
  3. Very specific:

    Catalog first: Check styles/ for a proven package that fits.

    line 89
  4. Distinctive:

    Each direction must be escapable-from-default: if a proposed direction is just the attractor with a new accent, cut it.

    line 92
  5. Very specific:

    Lock the direction: If it's a fresh synthesis worth reusing, save it as a style-reference in styles/<name>.md (the styles/cyberpunk.md …

    line 109
  6. Distinctive:

    Iterate the generator, re-render, re-look: Don't ship a frame you haven't looked at.

    line 177
  7. Distinctive:

    When a new design surprise appears, fold the fix back here — don't let the skill calcify.

    line 260

“Common” is measured: kept by at least 6 in 10 of the 174 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/design — commit to an art direction, then render and LOOK

This skill makes visual artifacts that win hearts, not "tasteful default" ones. It owns the part the generic loop keeps fumbling: the model is beautiful-by-default, which means bland-by-default. Asked to "make it nice" with no art direction, it regresses to the median of "tasteful product page" in its training — and that median is a specific look.

Core principle: Beauty that wins hearts is not "try harder" — it's a committed, named direction chosen BEFORE a line of HTML, an explicit ban on the default attractor, and a render→look→critique loop so the agent actually SEES what it made. Remove those three blockers and the same model that produced bland produces distinctive.

This is the identity counterpart to /diagram (which owns structure). For a diagram, the safe-polished default — rounded cards, hairline borders, dotted bg, one accent — is correct. For a hero/landing, that same default is the trap.


⛔ The default attractor — name it so you can escape it

Absent a committed direction, the model reaches for this exact combination. It is competent and forgettable. Recognize it and refuse it unless it is the chosen direction:

TellThe default (BAN unless chosen)
Backgroundwarm cream #f4f6f1 / off-white, with faint radial-gradient "orb" blobs + dot-grid texture
Accent palettesage green #1f6f54 + clay/terracotta #a25c1d (the "Claude/Anthropic" pairing)
Display typeFraunces (or another opsz serif) for headings + Inter body
Cardsborder-radius:18–22px, hairline #d9e0da border, soft 0 8px 26px rgba(...,.07) shadow
Moodcalm, editorial-SaaS, "thoughtful startup", safe

When you catch your draft using ≥2 of these rows and nobody asked for that look, you are in the attractor. Stop and commit a real direction.

Worked exemplar — the gap this skill closes. Same content, two outcomes. Bland (default attractor): --bg:#f4f6f1; --green:#1f6f54; --clay:#a25c1d, Fraunces headings, radius-18 cards, radial orbs. Reads "competent, generic, like every Claude page." Distinctive (committed direction — "cyberpunk control room"): --bg:#07090d near-black, neon --cyan:#2ef7ff / --magenta:#ff3df2 with color roles (cyan=system, magenta=AI), border-radius:0, clip-path corner notches, layered grid+scanline background, mono technical labels, a one-word glitch. Reads "mission systems online." The win came from the committed package, not a better model. The full package lives in styles/cyberpunk.md.


How to run

Stages in order. Hard gate: no HTML until a direction is locked (Stage 2). That single rule is what this skill buys you.

1. Brief (intake — keep it to these lines)

Settle before designing. Ask only what's missing from context; don't interrogate.

  • artifact — landing page / one-pager / hero section / pitch slide / theme-on-existing-markup
  • audience & job — who looks at it, what they should feel and do
  • content — the real copy/sections (use what exists; this skill does not rewrite IA/copy)
  • mood words — 2–4 adjectives the user actually wants (e.g. "tactical, premium, alive")
  • medium — screen only / also PDF-export / embeds in an existing app (inherits its tokens?)
  • hard constraints — brand colors that must appear, existing markup to theme, must-fit-on-one-screen
  • working path — where the draft lives, settled now: default .scratch/design/<slug>.html for exploration, or the target project's docs home for real work. Stages 3–6 (the brief's <path>, the render commands, the final landing) all use this one path.

2. Art-direction gate (the heart — interactive, never skip)

Before any drafting, propose 2–4 named, genuinely distinct directions. This is a fork with materially different outcomes → use the platform's multiple-choice question prompt so the user picks rather than free-types. Lead with your recommended direction (marked), offer real alternatives, always include "or your own".

Each direction is one tight block:

NAME — one-line mood ("cyberpunk control room: mission systems online, not a marketing hero")
palette: 5–7 tokens with ROLES (bg / surface / ink / muted + 1–2 accents, each accent = a job)
type:    display family + body family + label treatment (e.g. mono uppercase labels)
signature mechanic: the ONE move that makes it unmistakable (clip-path notches / brutalist
        hairlines / oversized editorial serif / glass + grain / terminal scanlines)
bans:    what this direction explicitly refuses (so it can't drift back to default)

Sourcing directions:

  • Catalog first. Check styles/ for a proven package that fits. A match → pull it whole.
  • Synthesize the rest. Riff named starting points from styles/INDEX.md, or invent. Make them different from each other — not three shades of the same idea.
  • Each direction must be escapable-from-default: if a proposed direction is just the attractor with a new accent, cut it.
  • Each direction must fit the AUDIENCE, not just be distinctive. A bold aesthetic that would intimidate or confuse the actual user (a neon/terminal console for a non-technical owner; a playful look for a compliance tool) is the wrong kind of distinctive. Name the audience in Stage 1 and weigh every direction against "would this person feel at home, or bounce at the door?" Distinctive ≠ appropriate.

Divergence mode (offer when the user is unsure or wants to compare): instead of picking from descriptions, render the top 2–3 directions as real comps — run Stage 3–4 (delegate

  • render + look) on each, in parallel — and let the user choose from actual PNGs. The binding skeptic (Stage 5) runs once, on the chosen comp after selection, not per-comp — you don't gate throwaways. This is how the cyberpunk-vs-CS comps were chosen. More tokens, far better calibration — use it when the look matters and the user can't decide blind. Give each comp its own slug (<slug>-cyberpunk, <slug>-swiss, …) so parallel renders don't collide on one path.

Lock the direction. If it's a fresh synthesis worth reusing, save it as a style-reference in styles/<name>.md (the styles/cyberpunk.md shape) so the next session inherits it.

3. Execute — delegate to the landing-designer persona

Design is artifact-shaped → delegate to a fresh subagent (CLAUDE.md §5). Inline drafting reads flat. Use the paste-ready brief below — it restates the persona inline (a fresh agent given only a path to a 13-line metadata card produces thin, un-embodied work; cold-read lesson). Fill every <bracket>; embed your locked direction verbatim.

You are a SENIOR CONVERSION-LANDING DESIGNER. You craft distinctive, high-converting
visual artifacts where aesthetics serve the goal. Operate to these standards:
 • Commit to ONE strong direction. A coherent token set (bg/surface/line/ink/accent)
   drives EVERY section — zero ad-hoc colors.
 • Aesthetics never cost the funnel: the primary action, proof, and any form/demo stay
   unmissable and readable.
 • Premium = restraint + craft (type scale, optical spacing, layered shadow, considered
   motion) — NOT effect-salad.
 • Hand-code self-contained HTML/CSS. No external fonts/images/CDNs unless the brief
   says so — system font stacks only. Theme via token overrides + a deliberate visual layer.
Anti-patterns you must avoid:
 • Default/templated look when a distinctive identity was asked for. SPECIFICALLY BANNED
   here: warm-cream bg, sage-green + clay accents, Fraunces serif, radial-orb blobs,
   dot-grid texture, soft 18px rounded cards — that is the attractor we are escaping.
 • Beautiful-but-illegible: low-contrast text, decorative type on body copy, motion that
   hides content.
 • Rewriting the copy or information architecture — only the VISUAL layer is in scope.

ART DIRECTION (locked — follow exactly, do not drift):
<paste the full locked direction block from Stage 2: name, mood, palette tokens with
 roles, type, signature mechanic, bans>

CONTENT (use verbatim, do not rewrite):
<the real sections/copy>

CONSTRAINTS: <brand colors / existing markup to theme / one-screen / PDF-export / 390px mobile, no horizontal overflow>
DELIVER: one self-contained .html file at <working path from Stage 1>. Verify contrast, and check
 mobile overflow at 390px via CDP device emulation (NOT a 390px render.sh window — see Stage 4)
 before returning. Tag any claim about existing markup [Observed: path:lines].

For a theme-on-existing-markup job, hand the agent the markup and have it deliver token overrides + a visual layer, not a rewrite.

4. Render → LOOK → fix (the loop that makes it good)

You can SEE the output — use it. Reuse the diagram pipeline; do not reinvent a renderer.

# render.sh <in.html> <out.png> [W] [H] [bg].  bg is RRGGBBAA (rgb + alpha LAST) under
# --headless=new — set it to the page background so any excess window space is seamless.
P=.scratch/design/<slug>          # the working path from Stage 1
.claude/skills/diagram/render.sh $P.html $P.png  1440 1800 07090dff   # desktop only — see the mobile note below
  • Read the desktop PNG. Critique against the locked direction, not against "is it nice": did it commit to the mechanic, or soften back toward default? Is the hierarchy unmissable? Contrast on body text?
  • Mobile: do NOT judge it from render.sh at a 390px width. render.sh runs under --force-device-scale-factor=2 and sizes the OS window, so a 390 request does NOT yield a true 390-CSS-px viewport — the @media breakpoint may not fire and you'll see a clipped desktop layout that fakes overflow (or hides real overflow). Check mobile via CDP device emulation: adapt the committed harness .claude/skills/sprite-anim/verify-anim.mjs (Emulation.setDeviceMetricsOverride with width:390, mobile:true), then compare documentElement.scrollWidth vs clientWidth (equal = no overflow) and screenshot. Keep a <meta name="viewport" content="width=device-width,initial-scale=1"> in the file or the breakpoints are moot on real devices.
  • Iterate the generator, re-render, re-look. Don't ship a frame you haven't looked at.
  • Render gotcha: render.sh screenshots the window viewport, not full page (see its header). For a tall page, set a tall H and iterate; or render section-by-section. Set the bg arg to the page background so excess space is seamless. @2x is already on.

5. Skeptic — design-failure-tuned, fresh, binding

A visual artifact is external-facing → run the skeptic (CLAUDE.md §6) as a fresh subagent with no drafting context. Tune it to how design fails — and judge fitness for the actual audience, not just craft. Put who the audience is in the brief so the skeptic can ask "would this person get it, or bounce?" Give it the rendered PNGs:

Verify this design artifact for its stated audience (<who they are>). Look at the attached desktop + 390px PNGs. Check:

  1. Did it commit? Or regress to the default attractor (cream/sage/clay/Fraunces/orbs/ soft-rounded)? Name any tell.
  2. First impression / intimidation — would the stated audience feel "this is for me, I get it", or "too technical/complex, I'm out"? Rate Low/Med/High and say why. A distinctive look that intimidates its own user is a FAIL, not a win — cool ≠ usable.
  3. Comprehension (5 sec) — can a first-timer tell what it is, what the key number/state means, and what (if anything) needs their action? What's confusing?
  4. Color encodes meaning, not just style — does a signal color mean ONE thing (red = the one action) or are loud colors sprayed decoratively so nothing reads as "this is what matters"? Decorative-only signal color is a defect.
  5. Hierarchy/funnel — primary action/message unmissable, or everything one flat weight? Does decoration ever bury content?
  6. Contrast — body text and labels readable (≥4.5:1 body, WCAG AA)? Any text lost on bg?
  7. Mobile — horizontal overflow at 390px (measure via CDP at a true 390 viewport, NOT a 390px render.sh window)? H1 readable? Stacking sane?
  8. Coherence + fidelity — every section on the same token set, no ad-hoc colors; matches the LOCKED direction; copy unchanged. Verdict: PASS / PASS-WITH-CHANGES / FAIL, with specific fixes. Never pass over an attractor regression, a contrast failure, OR a design that is distinctive-but-intimidating for its actual audience.

Verdict is binding. Never ship over red. (Skip the skeptic only for a throwaway the user is watching render in-loop.)

6. Ship

Paths first; state verified-vs-assumed; the artifact is turnkey (it actually opens and renders, mobile included). Land it in the target project's home, not .scratch/, unless it's an experiment. Record the run (tools/log-run.sh).


The style catalog (styles/)

Reusable art-direction packages — the same idea as personas/: reuse beats re-synthesize.

  • styles/INDEX.md — one line per direction + a roster of named starting points the gate can propose even before a full package exists.
  • styles/<name>.md — a full package (palette tokens with roles, background recipe, type + signature mechanic, component patterns, responsive + print rules, a reuse checklist). styles/cyberpunk.md is the proven seed; copy its shape.

A full package is earned by use: when a synthesized direction ships and is worth repeating, save it. Don't pre-fabricate a dozen thin entries.


Reach past this skill when

  • Structural diagram (architecture, data flow, lifecycle) → /diagram. There the safe-polished default is right; here it's the trap.
  • Pure copy / IA / messaging (no visual layer) → the doc/PRD flow. This skill themes given content; it does not write it.
  • Raster illustration / photographic hero → hand off a prompt (/diagram §"Reach past"): image-gen is UI-only here, no API. This skill is for hand-coded, exact, editable HTML/CSS.
  • Production component in a live app → the visual exploration can happen here, but the real change goes through the app's repo + frontend_engineer, never a one-off HTML file merged blind.

Files in this skill

FileWhat it is
styles/INDEX.mdcatalog index — proven packages + roster of named directions to propose
styles/cyberpunk.mdproven art-direction package (the worked exemplar's "good" side)
(renderer)reuse .claude/skills/diagram/render.sh in.html out.png [W] [H] [bg] — no own renderer

When a new design surprise appears, fold the fix back here — don't let the skill calcify.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.