Imported from alperien/blandaid (skills/blandaid-core/SKILL.md). Install upstream with npx skills add alperien/blandaid --skill blandaid-core. Copyright stays with the author (MIT).

blandaid core

The shared spine. Five generator skills and one detector skill reference this file so the vocabulary is defined once and cannot drift.

If you are writing a general design principle that is not specific to one pipeline, it belongs here. If you are reading a blandaid skill and it points at a section name below, this is the definition it means.

When this fires

Load alongside any blandaid skill. Load on its own when you need the shared vocabulary, the restraint rules, or the mode definitions without running a pipeline.

When this does not fire

This skill does not build anything. It has no phases and no output artifact. If the task is to produce a hero, a motion layer, reference images, a replication, or a restyle, load the matching generator skill and let it pull this one in.

The quality bar

"Premium" is not a mood. It is four measurable properties. A design that has all four reads as considered. A design missing any one reads as generated.

1. Hierarchy is unambiguous. A viewer's eye lands in one place first, and that place is the same for every viewer. Measured by squinting: blur the layout until text is unreadable. Exactly one element should still dominate. If three things compete, there is no hierarchy.

2. The type scale has real contrast. The ratio between the largest and smallest type on screen is at least 6 to 1. Machine defaults cluster at 2 to 1, which is why generated pages read flat. A 16px body with a 96px headline is a 6:1. A 16px body with a 32px headline is a 2:1 and looks like a template.

3. Space is uneven on purpose. Consistent spacing is a starting point, not a finish. Real layouts have one or two places where the space is much larger than the rhythm, and those gaps do structural work. Uniform padding everywhere is the signature of a component library used without judgment.

4. Every value traces to a decision. Any number in the output should have an answer to "why that one." A duration is 240ms because it is a small element moving a short distance. A color is that hex because it came from the reference or the brand. Values that cannot be justified were guessed, and guessed values cluster around the statistical mean. "Provenance", below, defines which answers count.

Provenance

Every non-trivial visual decision names a source, and the source sits outside this skill set.

Sources that count:

• The subject. What the product is, what it touches, where it lives. A Postgres tool lives in terminals, CI logs, and branch diagrams. A bakery lives in flour, ovens, and handwritten labels.
• The brand: tokens, voice, and material the user already owns.
• The content: what the words and images on this page actually say and need.
• The medium: print-length reading, a dense operator screen, a phone outdoors.
• A measured constraint: a contrast floor, a performance budget, an input method.
• The user's stated taste, recorded in their own words.

Sources that do not count:

• A rule in this skill set. Rules veto options; they never nominate them. "Core bans pure black" explains why not #000000 and says nothing about which ink this subject wants instead.
• What other tools are producing, or "what good sites look like right now."
• "It reads as premium." Premium is a verdict, not a source.

The test: take the five most visible decisions (display face, palette, texture, layout grammar, motion character) and write one line each naming the source. A line that cites a rule or a trend means the decision was defaulted, and the output will match every other page defaulted the same way.

The failure that prompted this section: a Postgres branching tool styled as an editorial magazine. Serif "because pairing," sage "because restraint," grain "because atmosphere." Every answer cited a rule. The same stylesheet would have fit a fragrance brand without edits, which is the proof that it fit neither.

Pick a register

Before any visual work, name three candidate registers derived from the subject, with one line on why each fits. Pick one and commit. In direct mode, show the candidates and the choice.

A register is the kind of printed or built thing the page resembles: a terminal, a spec sheet, a field guide, a magazine, a poster wall, a lab notebook, an appliance manual, a transit map. The list is open. Most subjects point somewhere specific, and it is rarely where the previous subject pointed.

Editorial (display serif, italic emphasis, paper tones, kickers) is one register among many, and it is never the default. Pick it only when the subject is genuinely editorial: a publication, longform writing, a literary brand. As of mid 2026 it is also the register machine output converges on, so choosing it without subject provenance is indistinguishable from not choosing at all. The detector catalogs that cluster as tells 33 to 40.

Authorship

Passing every rule in this set proves the absence of known defects. It does not put a person in the output. Two requirements do.

One rubric-breaking decision. At least one visible choice that this skill set would flag, kept on purpose, defended in one line in a code comment. A page in full compliance with a published rubric reads as what it is: rubric output. The break is where a viewer senses that someone decided.

Uneven investment. Name the one moment that gets disproportionate craft, and spend there. Let at least one region stay plain: a footer that is two lines, a table that is just a table. Checklists produce even polish. People spend attention where the argument lives and stop when it stops mattering.

Neither requirement is decoration. Fake roughness applied evenly is polish with extra steps. The break and the spike need sources like everything else; the difference is that their source is your judgment about this subject, stated out loud.

Shared easing vocabulary

Five curves cover almost everything. Use these names. Skills that need specialized curves extend this set rather than redefining it.

:root {
  /* Default for entrances and most UI motion. Fast start, long settle. */
  --ease-out: cubic-bezier(0.16, 1, 0.3, 1);

  /* Tighter and more responsive. Hovers, toggles, small state changes. */
  --ease-snap: cubic-bezier(0.22, 1, 0.36, 1);

  /* Slight overshoot. Use where a element should feel physical. */
  --ease-spring: cubic-bezier(0.34, 1.56, 0.64, 1);

  /* Slow in and out. Full-screen transitions and large travel only. */
  --ease-dramatic: cubic-bezier(0.65, 0, 0.35, 1);

  /* Symmetric. Loops, ambient motion, anything that repeats. */
  --ease-in-out: cubic-bezier(0.05, 0.7, 0.1, 1.0);
}

The CSS keywords ease, ease-in, ease-out, ease-in-out and linear are banned in generated output, with one exception: linear is correct for a constant-rate loop such as a marquee or a spinner.

The keywords are banned because they are the browser default and the model default at the same time. ease is cubic-bezier(0.25, 0.1, 0.25, 1), a curve that starts slow. Motion that starts slow reads as lag.

Durations, in the same spirit:

:root {
  --dur-instant: 120ms;  /* color and opacity on small elements */
  --dur-fast:    240ms;  /* buttons, toggles, hovers */
  --dur-base:    400ms;  /* cards, panels, most entrances */
  --dur-slow:    700ms;  /* section reveals, large travel */
  --dur-scene:  1200ms;  /* full-viewport transitions only */
}

Distance sets duration. A 4px shift at 700ms looks broken. A full-screen wipe at 120ms is invisible.

Shared spatial vocabulary

Type scale. Pick a ratio and hold it. 1.25 for dense interfaces, 1.333 for marketing pages, 1.5 or higher for editorial and hero work. Generate the scale from the ratio rather than choosing sizes by eye. The largest step should clear the 6:1 contrast bar in "The quality bar".

Fluid sizing between breakpoints:

h1 { font-size: clamp(2.75rem, 8vw + 1rem, 9rem); }

Spacing scale. A 4px base with a non-linear progression: 4, 8, 12, 16, 24, 32, 48, 64, 96, 128, 192, 256. Values off this scale need a reason.

Section rhythm. Vertical padding between major sections should be at least clamp(6rem, 12vw, 14rem). The most common spatial failure in generated layouts is sections that are too close together, which makes a page read as one undifferentiated column.

Measure. Body copy sits between 60 and 75 characters per line. Use ch units, not percentages.

Grid. A 12 column grid is the default. Using all 12 columns for every element is not a grid, it is a stack. At least one element per viewport should break the expected column span.

Palette discipline

Count. Two or three hues total, plus neutrals. A fourth hue needs a justification. Generated palettes sprawl because every new component brings its own color.

Neutrals are never pure. #000000 and #FFFFFF are the two most common tells in generated work. Real designs use off-black and off-white with a temperature bias that matches the accent.

/* Bad */
--bg: #FFFFFF;
--fg: #000000;

/* Good, warm bias */
--bg: #FAF8F5;
--fg: #14110E;

Accent ratio. The accent color covers under 10 percent of the visible surface. An accent applied to every button, link, badge and icon stops being an accent.

Contrast is a floor, not a target. Meet WCAG AA at minimum. Meeting it exactly on body text is a sign the palette was checked rather than designed.

A supplied brand palette outranks every rule in this section. See "Universal restraint".

Universal restraint

Every blandaid skill is deliberately aggressive. Aggression without a stop condition destroys good work. These rules outrank the aesthetic rules in every skill.

Signals to preserve. When you see these in the input, they are decisions, not defects:

• A brand palette, brand font, or design system named by the user or present in the codebase. You do not have standing to override a brand.
• A deliberately quiet type scale where the content is dense and the reading task is long. Documentation and reference material are supposed to be calm.
• High-contrast or large-text choices that look heavy-handed. They are usually accessibility requirements.
• Dense tables and data grids. Density is the feature. Adding whitespace to a trading screen or an admin table makes it worse.
• Existing motion tokens in a codebase. Matching them beats importing ours.
• Content that is genuinely short. A one-sentence page does not need six sections. Do not generate filler to fill a layout.

Contexts where the blandaid aesthetic is wrong:

• Internal tools, admin panels and dashboards, where speed of use beats impression.
• Government, healthcare, legal and compliance interfaces, where convention is a usability feature and novelty is a risk.
• Documentation sites and API references.
• Any product where the user named an existing design system such as Material, Carbon, Polaris or an internal one.

What to do when restraint wins. Say so, in one sentence, and explain the call. Then do the smaller version of the job. Silently skipping a skill and silently maximalizing are both failures. The user should always know which mode you chose and why.

Modes

Every blandaid skill supports these three. The names mean the same thing everywhere.

direct. The user invoked the skill as the task. Run the full pipeline and show the work: the extraction, the gates, the diff table. The process output is part of the deliverable.

embedded. Another skill or a larger job is using this as one step. Run the same pipeline internally and output only the artifact. No extraction sheets, no gate tables, no commentary.

audit. Point the skill at existing output and report only. Change nothing. Return the diff table with PASS and FAIL rows, then list the highest-severity failures.

Default to direct unless the calling context makes it obvious you are one step inside a larger job.

The detector

skills/design-humanizer/SKILL.md holds the catalog of AI design tells and the audit loop. Every generator skill in this set ends with a design-humanizer pass in audit mode.

One exception: pixel-perfect replicates a reference exactly. If the reference contains a tell, the tell is the specification. There, design-humanizer labels rather than corrects.

The detector's tells 33 to 40 are dated: they name the register machine output defaults to as of mid 2026. When a page avoids the first 32 tells by relocating there, "Provenance" and "Pick a register" above are the counterweight.