Ground a product's UI/UX in established HCI and design theory instead of gut feel. Point it at a product (an idea, a spec, or a running UI); it works out which theoretical lenses apply, then produces concrete design suggestions, a design guideline to build against, and a severity-scored health-check table that audits whether the current design meets human expectations.
"Overview first, zoom and filter, then details-on-demand." — Ben Shneiderman, "The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations" (1996)
Two roles. Invoke it directly to generate a guideline for a new product or run a
health-check on an existing one; or, as a library, other skills read
references/theory-catalog.md to pull the right lens for
a design decision (e.g. /aep-model reaching for the information-seeking and
graphical-perception lenses when shaping a data-heavy screen).
When to Use — and Its Boundaries
Reach for this skill to produce a design guideline before agents build, to run a usability / heuristic health-check on a drafted or running UI, or when a design decision would benefit from the relevant established theory (which chart encoding, how to structure exploration, how to make an agent's actions legible).
Its boundary against neighboring skills — stated once, here:
- Taste decisions (palette, voice, journey feel) belong to
/aep-calibrate. This skill supplies the theory a calibration leans on; it does not decide taste. - Object / screen IA (which objects, what fields, which screens) belongs to
/aep-model. - Factual correctness of artifacts belongs to
/aep-validate; the heuristic health-check is its advisory design-quality complement. - It does not replace real user testing — heuristics predict where users will struggle; they do not observe actual users. Treat the health-check as evidence, not proof.
A backend-only surface with no human-perceived UI is out of scope — but a CLI that wraps an agent still benefits from family E (Gulf of Envisioning), and any terminal output is a human-perceived UI for family G (color-alone signals, contrast).
Output goes to the conversation by default, or to a standalone markdown doc when the
user asks to save it — it never writes to product-context.yaml or any AEP schema file.
This is an advisory leaf skill that adds no taxonomy.
The Method
Product-agnostic, because it classifies tasks and data, not a product type. Full step
detail, the lens-selection rules, the worked examples, the live-UI evidence-gathering
protocol, and the output templates are in
references/method-and-templates.md.
Two depths. A quick check scores only the Baseline Ten — a distilled
cross-family checklist
(method-and-templates.md#the-baseline-ten-quick-check) —
and answers "is this OK?" in minutes; default to it for a lightweight question ("is this
good UX?"). A deep audit runs all seven steps below; default to it for a guideline or
review request. Escalation rule: any major/catastrophe finding (severity 3–4) upgrades
a quick check to a deep audit, so speed never trades away rigor.
- Characterize — what is the product, who uses it, what are they trying to do? If a running UI or spec exists, point to it.
- Abstract — classify the primary user tasks (monitor / compare / look-up / diagnose / explore / decide / create / reconstruct-provenance / converse) and the data/content type. (Munzner's domain → abstraction move.)
- Select lenses — pick the applicable theory families from the catalog based on the task+data profile; families A, B, F, G always fire, C/D/E on triggers (selection rules in the reference). Note why each family fired.
- Synthesize — per selected lens, turn its Apply line into ≥1 concrete,
product-specific suggestion, each tagged
(→ lens id)back to its theory. Done when every selected lens carries at least one traced suggestion. - Emit a design guideline — prescriptive, grouped by family: "the product should…".
- Emit a health-check table — one row per checkable expectation, each traced to a theory and carrying a Nielsen 0–4 severity, producing a scored "meets human expectations?" audit.
- (Opt-in) persist — on request, write the report to a standalone markdown file
(default
docs/design-review/<scope>-<date>.md).
The Theory Catalog
Grouped into families so lenses can be selected per product. Full entries — each with
core question · when it applies · how to apply · what to check · source — are in
references/theory-catalog.md, which is append-only and
extensible (it documents how to add a theory).
| Family | Lenses (summary) |
|---|---|
| A. Usability & interaction heuristics | Nielsen's 10 · Shneiderman's Eight Golden Rules · Norman's principles + Gulfs + Seven Stages · Jakob's / Tesler's / Postel's laws |
| B. Cognitive & perceptual laws | Cognitive Load · Miller's · Hick's · Fitts's · Doherty Threshold · Gestalt · preattentive attributes · Aesthetic-Usability · memory/motivation biases |
| C. Information seeking & navigation | Shneiderman's Visual Information-Seeking Mantra · Information Foraging (scent) · Focus+Context / multiscale trio · IA / wayfinding |
| D. Data visualization & encoding | Munzner's Nested Model + task typology · Graphical Perception · Grammar of Graphics / Bertin · Heer & Shneiderman · Tufte |
| E. Human-AI & agent-specific UX | Gulf of Envisioning · Microsoft's 18 HAI Guidelines · Google PAIR · trust calibration · observability · steerability/reversibility · Grice's maxims · CASA |
| F. Evaluation & process methods | Heuristic Evaluation · Cognitive Walkthrough · Nielsen 0–4 severity scale · GOMS/KLM · Norman's stages-of-action as a lens |
| G. Accessibility & inclusive design | WCAG 2.2 POUR · keyboard & focus · contrast & legibility (never color alone) · inclusive design (permanent/temporary/situational impairments) |
How It Fits AEP
A design-lens pass feeds forward; control then returns to the calling context.
- Guideline for a new UI →
/aep-modelturns it into object/screen structure (drawing families C and D), and/aep-calibrateux-flow/visual-designcaptures the human's taste decision (drawing A, B, C, E) — the report seeds the brief and pre-audits the flow before the human calibrates. - Health-check on a built UI → route findings through
/aep-reflect(a "works vs right" gap is a calibration or refinement item), or fix directly if it's a clear defect./aep-validatepoints here for its advisory design-quality pass (family F). - Recurring audit → pair with
/loopto re-run the health-check each layer.
Why it's shaped this way
A product-agnostic task+data method (not a product-type lookup), a leaf skill beside
/aep-model and /aep-calibrate rather than inside them, hybrid output, Nielsen 0–4
severity, two depths, and always-on accessibility — the reasoning is recorded in
docs/decisions/design-lens-rationale.md.