Empirical design principles
When to use
- Deciding spacing, grouping, hierarchy or emphasis on a page, before it is built.
- Reviewing a layout where a principle is being cited as the reason it works.
- Someone proposes a rule of thumb — "seven items max", "reduce the options" —
and it needs checking against what the underlying work actually says.
When NOT to use
- Reaching for a usability heuristic rather than an empirical principle.
Nielsen's ten, Shneiderman's eight and the rest are not here, deliberately.
They make a different kind of claim, so they fail differently: a heuristic is
accumulated judgement organised for inspection, and it fails by missing a
problem. A principle here predicts, and fails by predicting wrongly. Applying
one is an act of judgement you defend; applying the other produces a number you
test. That absence is declared, not covered —
see usability-heuristics.
- Reaching for Norman's constraints, forcing functions, or natural
mapping. Neither skill covers them. That one is a real gap, not a declared
absence: cite the edition and page yourself — DOET rev. ed. 2013 ch. 4 for
constraints and forcing functions, ch. 1 and 3 for mapping — and do not read
either file's silence as clearance. The two gulfs are covered, in
composing-a-screen.
- Picking token values. Contrast, type scale, spacing steps and control heights
are owned by the system, not by the findings here — see Cross-references.
- Judging craft quality. Whether something is elite is a separate question
from whether its arrangement is well-formed.
They are descriptive, not prescriptive
Every principle here describes what perception or behaviour does under stated
conditions. None instructs you what to build. They are ceteris paribus claims —
true when nothing else varies — and an interface is precisely the case where
everything else varies. Palmer, who introduced common region, says so himself:
he calls them "principles" or "factors" rather than laws because "they lack the
quantitative structure of standard scientific 'laws'."
So use them forward, never backward. Forward: predict what a layout will do,
then build it and check. Backward: ship the layout, then name a principle that
sounds like what you did. The second is the dominant failure and it is invisible,
because a post-hoc citation is always available for any arrangement — a card grid
with 12/16/20px spacing gets defended as "grouping by proximity" when applied
forward the same principle predicts it fails. Used before, a principle produces
a number you can test; used after, only a name for what already exists.
The one that changes how you space things
Proximity works on ratios, and is scale-invariant. Grouping strength follows
relative distance, not absolute distance — the same layout scaled up groups
identically. Below roughly a 1.5 ratio between competing gaps, no grouping
reliably dominates.
That refutes the common framing. "8px means related, 24px means separate" works
because it is 3:1, not because of the numbers, and it silently stops working the
moment a neighbouring group uses a different base. A spacing scale that
multiplies across breakpoints preserves grouping; one that adds a fixed offset
("+8px on desktop") compresses every ratio toward 1 and dissolves the structure it
was meant to protect.
Check ratios, not pixels. 24 against 32 is 1.33 — near-bistable, and readers will
parse it differently on different visits.
Two that hold, inside a stated scope
- Fitts's — movement time rises with distance and falls with target size,
logarithmically. It models movement to a target of known location, not
finding one, and its edge corollary needs a pointer to pin against, not a
finger.
- Jakob's — users arrive with a model built on other sites, so an unfamiliar
convention costs errors on first encounter. The measured support is control
placement and behaviour only; the visual half is Nielsen's opinion.
Four that do not say what design writing says
Read interaction-laws before citing any of these.
- Hick's Law — "reduce the number of options and nest them into categories"
is not supported, and on the law's own maths is backwards: nesting charges an
extra intercept at every level, so the interface looks simpler and measures
slower.
- Miller's Law — "7±2 items in a menu" is the most repeated misquote in
interface design. The paper is about immediate memory for unrelated items and
supports no menu limit in any form.
- The aesthetic-usability effect (Kurosu & Kashimura 1995; replicated with a
controlled manipulation by Tractinsky, Katz & Ikar 2000) — users rate
attractive designs as more usable. It is a claim about perception and
tolerance. The study that manipulated real usability found aesthetics did not
move actual performance, so it never licenses deferring usability work. It is
not Nielsen's heuristic 8, "aesthetic and minimalist design" — different
authors, and a density claim about irrelevant information competing with
relevant, saying nothing about beauty. Fusing the two is the commonest error
on this entry; the heuristic side is in
usability-heuristics.
- Von Restorff — a memory effect, measured by delayed recall of lists.
Nothing in it addresses click-through, so it cannot justify "make the CTA
orange and it will be clicked."
Verification
Before citing a principle in a review, state the prediction it makes and the
observation that would falsify it. If neither is available, the citation is
decoration and the finding should stand on what was actually observed.
For spacing claims, compute the ratio between competing gaps and quote it.
Cross-references
- usability-heuristics — the judgement half:
Nielsen's ten and heuristic evaluation, Norman by edition, Shneiderman's eight,
progressive disclosure, recognition over recall.
- reviewing-design-work — the dispatcher that
routes here when a finding is being justified by naming a principle.
- visual-polish — the rendered-surface lens that catches
two gaps reading identical; it defers to the ratio above for how far apart they
then have to be, so do not answer that question from the render alone.
- designing-elite-ui — the craft bar, once the
layout is sound.
- spacing-system — the spacing ramp these ratios
are expressed in.
- component-sizing-principles — control
heights and the 24 CSS px interactive floor.
- apca-contrast — figure/ground separation as a
measured contrast obligation rather than an intuition.
- consuming-a-design-system, plus your
own project's design-system skill if it has one — this skill settles the
arrangement, the system settles the value.
Detail sits beside this file on disk — npx skills add copies references/
wholesale — so read it there, not over the network:
gestalt ·
interaction-laws ·
convention-and-memory.
No blob/main URL is given: that promises a path that never moves.
Sources
The three reference files linked above hold 78 source entries. Verification is
partial and the scope is the claim: about 20 were opened and read against the
original — those marked [read in full], plus the ones the inline corrections
turn on — and the rest are DOI- or URL-resolved only. A resolving DOI shows a
work exists, not that it says what the citing line claims. Corrections are
recorded inline, including who actually established figure/ground, common region
and element connectedness, none of whom is Wertheimer.
1---2version: "1.0.0"3digest: "9b90fefae38b" # is your copy current? github.com/thrillmade/agent-skills/blob/main/docs/skill-versions.json4origin: "https://github.com/thrillmade/agent-skills"5name: empirical-design-principles6description: Use when a layout decision is about to be defended by naming a principle — Gestalt grouping and common region, Fitts's, Hick's, Miller's, Jakob's, von Restorff, the aesthetic-usability effect — or when a rule of thumb like "seven items max" needs checking against what the underlying work actually says. Covers the empirical tradition only, meaning principles that make a prediction which can be tested and can fail. Says which ones predict something real, which four do not say what design writing claims, and how to use one forward as a prediction rather than backward as a post-hoc justification. Does NOT cover usability heuristics — Nielsen's ten, Norman's signifiers, Shneiderman's eight, progressive disclosure, recognition over recall — which are design judgement organised for inspection, not falsifiable prediction; those are in usability-heuristics. Norman's constraints, forcing functions and natural mapping are in neither skill and covered nowhere yet; the two gulfs of execution and evaluation are covered in composing-a-screen.7metadata:8 source: https://skillmd.com/skills/thrillmade/empirical-design-principles9---1011# Empirical design principles1213## When to use1415- Deciding spacing, grouping, hierarchy or emphasis on a page, before it is built.16- Reviewing a layout where a principle is being cited as the reason it works.17- Someone proposes a rule of thumb — "seven items max", "reduce the options" —18 and it needs checking against what the underlying work actually says.1920## When NOT to use2122- **Reaching for a usability heuristic rather than an empirical principle.**23 Nielsen's ten, Shneiderman's eight and the rest are not here, deliberately.24 They make a different *kind* of claim, so they fail differently: a heuristic is25 accumulated judgement organised for inspection, and it fails by missing a26 problem. A principle here predicts, and fails by predicting wrongly. Applying27 one is an act of judgement you defend; applying the other produces a number you28 test. That absence is declared, not covered —29 see [usability-heuristics](../usability-heuristics/SKILL.md).30- **Reaching for Norman's constraints, forcing functions, or natural31 mapping.** Neither skill covers them. That one is a real gap, not a declared32 absence: cite the edition and page yourself — DOET rev. ed. 2013 ch. 4 for33 constraints and forcing functions, ch. 1 and 3 for mapping — and do not read34 either file's silence as clearance. The two gulfs *are* covered, in35 [composing-a-screen](../composing-a-screen/SKILL.md).36- Picking token values. Contrast, type scale, spacing steps and control heights37 are owned by the system, not by the findings here — see Cross-references.38- Judging craft quality. Whether something is *elite* is a separate question39 from whether its arrangement is well-formed.4041## They are descriptive, not prescriptive4243Every principle here describes what perception or behaviour *does* under stated44conditions. None instructs you what to build. They are *ceteris paribus* claims —45true when nothing else varies — and an interface is precisely the case where46everything else varies. Palmer, who introduced common region, says so himself:47he calls them "principles" or "factors" rather than laws because "they lack the48quantitative structure of standard scientific 'laws'."4950**So use them forward, never backward.** Forward: predict what a layout will do,51then build it and check. Backward: ship the layout, then name a principle that52sounds like what you did. The second is the dominant failure and it is invisible,53because a post-hoc citation is always available for any arrangement — a card grid54with 12/16/20px spacing gets defended as "grouping by proximity" when applied55forward the same principle predicts it fails. Used before, a principle produces56a number you can test; used after, only a name for what already exists.5758## The one that changes how you space things5960**Proximity works on ratios, and is scale-invariant.** Grouping strength follows61relative distance, not absolute distance — the same layout scaled up groups62identically. Below roughly a 1.5 ratio between competing gaps, no grouping63reliably dominates.6465That refutes the common framing. "8px means related, 24px means separate" works66because it is 3:1, not because of the numbers, and it silently stops working the67moment a neighbouring group uses a different base. A spacing scale that68*multiplies* across breakpoints preserves grouping; one that adds a fixed offset69("+8px on desktop") compresses every ratio toward 1 and dissolves the structure it70was meant to protect.7172Check ratios, not pixels. 24 against 32 is 1.33 — near-bistable, and readers will73parse it differently on different visits.7475## Two that hold, inside a stated scope7677- **Fitts's** — movement time rises with distance and falls with target size,78 logarithmically. It models movement to a target of *known* location, not79 finding one, and its edge corollary needs a pointer to pin against, not a80 finger.81- **Jakob's** — users arrive with a model built on other sites, so an unfamiliar82 convention costs errors on first encounter. The measured support is control83 placement and behaviour only; the visual half is Nielsen's opinion.8485## Four that do not say what design writing says8687Read [interaction-laws](references/interaction-laws.md) before citing any of these.8889- **Hick's Law** — "reduce the number of options and nest them into categories"90 is not supported, and on the law's own maths is backwards: nesting charges an91 extra intercept at every level, so the interface looks simpler and measures92 slower.93- **Miller's Law** — "7±2 items in a menu" is the most repeated misquote in94 interface design. The paper is about immediate memory for unrelated items and95 supports no menu limit in any form.96- **The aesthetic-usability effect** (Kurosu & Kashimura 1995; replicated with a97 controlled manipulation by Tractinsky, Katz & Ikar 2000) — users *rate*98 attractive designs as more usable. It is a claim about perception and99 tolerance. The study that manipulated real usability found aesthetics did not100 move actual performance, so it never licenses deferring usability work. It is101 **not** Nielsen's heuristic 8, "aesthetic and minimalist design" — different102 authors, and a density claim about irrelevant information competing with103 relevant, saying nothing about beauty. Fusing the two is the commonest error104 on this entry; the heuristic side is in105 [usability-heuristics](../usability-heuristics/SKILL.md).106- **Von Restorff** — a memory effect, measured by delayed recall of lists.107 Nothing in it addresses click-through, so it cannot justify "make the CTA108 orange and it will be clicked."109110## Verification111112Before citing a principle in a review, state the prediction it makes and the113observation that would falsify it. If neither is available, the citation is114decoration and the finding should stand on what was actually observed.115116For spacing claims, compute the ratio between competing gaps and quote it.117118## Cross-references119120- [usability-heuristics](../usability-heuristics/SKILL.md) — the judgement half:121 Nielsen's ten and heuristic evaluation, Norman by edition, Shneiderman's eight,122 progressive disclosure, recognition over recall.123- [reviewing-design-work](../reviewing-design-work/SKILL.md) — the dispatcher that124 routes here when a finding is being justified by naming a principle.125- [visual-polish](../visual-polish/SKILL.md) — the rendered-surface lens that catches126 two gaps reading identical; it defers to the ratio above for how far apart they127 then have to be, so do not answer that question from the render alone.128- [designing-elite-ui](../designing-elite-ui/SKILL.md) — the craft bar, once the129 layout is sound.130- [spacing-system](../spacing-system/SKILL.md) — the spacing ramp these ratios131 are expressed in.132- [component-sizing-principles](../component-sizing-principles/SKILL.md) — control133 heights and the 24 CSS px interactive floor.134- [apca-contrast](../apca-contrast/SKILL.md) — figure/ground separation as a135 measured contrast obligation rather than an intuition.136- [consuming-a-design-system](../consuming-a-design-system/SKILL.md), plus your137 own project's design-system skill if it has one — this skill settles the138 arrangement, the system settles the value.139140Detail sits beside this file on disk — `npx skills add` copies `references/`141wholesale — so read it there, not over the network:142[gestalt](references/gestalt.md) ·143[interaction-laws](references/interaction-laws.md) ·144[convention-and-memory](references/convention-and-memory.md).145No `blob/main` URL is given: that promises a path that never moves.146147## Sources148149The three reference files linked above hold 78 source entries. Verification is150partial and the scope is the claim: about 20 were opened and read against the151original — those marked `[read in full]`, plus the ones the inline corrections152turn on — and the rest are DOI- or URL-resolved only. A resolving DOI shows a153work exists, not that it says what the citing line claims. Corrections are154recorded inline, including who actually established figure/ground, common region155and element connectedness, none of whom is Wertheimer.