deterministic-design – Design & UI skill for AI agents

When to Use

Use to catch AI-generated UI that "looks off", is misaligned or centered-mush, or fails usability — when you need reproducible spatial measurements and an evidence-based usability review. Metrics do not prove user success, accessibility or overall design quality. Compose it with any taste/token design skill before reporting design "done".

_Source: [connerkward/deterministic-design-skill](https://github.com/connerkward/deterministic-design-skill) (MIT)._

deterministic-design

Thesis: determinism beats AI randomness. A model can't trust its own eye on layout — so

don't. Render the UI and *measure* it.

Two sub-skills (load as needed):

• [design-spatial](https://github.com/connerkward/deterministic-design-skill/blob/main/design-spatial/SKILL.md) — deterministic layout audit: explicit grid

+ 8-pt spacing, and layout-audit.js computes centroid / optical-center / pixel-oracle

balance and draws an annotated screenshot. Numbers, not vibes. Plus a render-then-

critique vision loop.

• [design-ux](https://github.com/connerkward/deterministic-design-skill/blob/main/design-ux/SKILL.md) — usability audit: scores the rendered UI against

Nielsen's 10 + interaction heuristics via an authorized separate judge or clearly disclosed self-review → prioritized fix list.

This improves existing design skills (including the default Anthropic one) by adding the

layer they lack — it doesn't just advise on taste, it renders, measures, and judges the

output. Composable with any design skill.

In central this lives as a subdir of ckw-design; it publishes separately as

deterministic-design-skill (its own distribution) via publish-skill. One of the two

flagship narratives — the *determinism* one; its sibling is human-in-the-loop (lookdev).

Example

User request:

> Review this interface with @deterministic-design, identify the highest-impact design problems, and propose an implementation-ready improvement.

Limitations

• Layout metrics and vision-judged audits catch many spatial and usability failures, but they are not a substitute for product judgment or user testing.
• The workflow requires a rendered UI or screenshot; record viewport, scale, fonts, capture state and measurement definitions. Separate layout measurements, subjective vision judgments and actual user/accessibility tests.
• The linked sub-skills and layout-audit.js are external, not bundled. Read and pin them before use; the link is not executable availability. Delegation requires existing explicit authorization, and screenshots may contain private data.
• Automated scoring can miss brand nuance, copy tone, accessibility needs, and domain-specific user expectations.