AI DESIGN LAB

AI Design
Lab.

Tools, prompts, and workflows I actually use.
Updated as things change.

What's in active use.

These are the tools earning their place in the workflow right now. Not a wishlist, a working stack.

Tool Role in workflow Verdict
Claude (Sonnet) Brief synthesis, copy, code review Daily driver
Claude Code Component scaffolding, refactoring Replaces most Cursor use
Gemini Generative concepting, image prompts Early ideation only
Google Stitch Design system tokens → UI Figma complement
Cursor In-editor AI completions Used selectively

Prompts that earn consistent results.

Tested in real project work. The format matters as much as the content.

Brand brief synthesis

Claude (Sonnet)

Turn a raw creative brief into structured brand direction

You are a senior brand strategist. I'll give you a raw brief. Extract and return:
1. Core brand tension (one sentence: what the brand is navigating between)
2. Three personality adjectives with their opposites (e.g. "Calm / not clinical")
3. One-sentence positioning statement
4. Two visual metaphors with rationale

Be specific, not generic. Reject adjectives like "innovative" and "trusted" unless I've given you evidence for them. Here's the brief: [PASTE BRIEF]

Token hierarchy from brand decisions

Claude Code

Map brand decisions to a CSS custom property architecture

Given this brand brief: [BRIEF], generate a CSS custom properties token hierarchy.

Output structure:
- 6 base colour tokens with semantic names and suggested hex values
- 3 typographic scale steps with sizes and intended use
- 4 spatial tokens (xs, sm, md, lg) with values in rem

Format as valid CSS custom properties inside :root {}. Do not invent values I haven't implied in the brief. Flag any gaps explicitly.

React component scaffold

Claude Code

Get a working component from a token-aware design spec

Scaffold a React functional component for [COMPONENT NAME].

Design spec: [DESCRIPTION OF VISUAL BEHAVIOR]
Token constraints: uses --ink, --paper, --accent, --bone, --rule from our CSS custom properties.
Requirements: named export, TypeScript interfaces defined inline, no inline styles, no hardcoded colours or spacing, no magic numbers.

Do not add animation or interaction beyond what I've described. Do not add props I haven't asked for.

Design critique before shipping

Claude (Sonnet)

Get a critical review of visual work before client presentation

You are a design director reviewing work before a client presentation. Critique this design: [DESCRIBE THE DESIGN OR PASTE A SCREENSHOT].

Check for:
- Visual hierarchy (is the reading order clear and intentional?)
- Colour contrast (any accessibility failures?)
- Spacing consistency (does it follow a system or is it arbitrary?)
- Typography (are weight and size decisions doing semantic work?)
- Alignment (is the grid respected?)

Rank issues P1 (blocks shipping), P2 (should fix), P3 (polish). No encouragement. No filler. Just problems and specific fixes.

Technical diagram negative prompt

Midjourney / Flux

Strip AI failure modes from architecture and flow diagrams

photorealistic, stock photography, human faces, hands, office environment, 3D render, chrome surfaces, neon glow, lens flare, gradient sphere, generic icons, marketing cliché, mixed typefaces, serif body text, blurry elements, low contrast, watermark, decorative borders, drop shadows on flat elements, perspective distortion, isometric inconsistency

What was tested, what held.

A changelog of AI workflow experiments: what worked, what didn't, and what's been retired.

2026-05 active

Claude Code for component scaffolding

Replaced ~70% of Cursor component work. Output quality higher when full token context is provided upfront.

2026-04 selective

Google Stitch for design system scaffolding

Useful for rapid component exploration; outputs require significant cleanup before production use.

2026-04 active

Gemini for moodboard generation

Generated 120 direction boards in parallel; 12 curated, 3 selected. Faster than Pinterest for early ideation phase.

2026-03 active

Multi-model pipeline for brand consistency

Claude for synthesis, Gemini for generation, ChatGPT for copy alternatives. Maintained consistency across AI outputs by routing tasks by model strength.

2026-02 active

AI-assisted competitive research

3-day research compressed to 4 hours. Useful for pattern recognition; human judgment still required for "which pattern matters here."

2026-01 retired

Midjourney for technical diagram generation

Better for abstract direction boards than technical diagrams. Diagram generation still requires manual SVG work for precision.

Interested in AI-assisted design ops for your team?

I consult on AI workflow integration for design and product teams: tools, systems, and the judgment layer that makes it work.

Get in touch →
WhatsApp Call Email