66666-design/ai-design-director

AI Design Director (ai-design-director)

Based on the ideas from: Anshu Chimala's "How to turn your AI into a world-class designer" (Lenny's Newsletter, 2026-09-01). This document is an independent rewrite and extension, not containing the original text.

Why Default Outputs Always Feel "AI-Made"

Large models always choose the next word that is "most likely to satisfy everyone." As a result, design decisions (colors, layout, typography) become averaged: purple gradients, left-text-right-image, top navigation, centered big button. Just saying "make it unique" or "randomize it" won't work—the model will write something that "sounds random" but ends up being the same every time. Variation must be injected from outside the model, and taste must be guarded by a human or an independent reviewer.

Overall Process

1. Discover: First diverge, find directions that others can't.
2. Define: Deepen the direction, form your own design identity.
3. Deliver: Polish, subtract, remove AI traces.

Each stage must produce something visible (screenshots), not rely only on text-based self-evaluation.

---

I. Discover

Technique 1: Random Seed String (String Seed of Thought)

The model itself cannot truly randomize, so use shell to generate a random string and ask the model to "read" creative directions from it.

python3 scripts/seed.py         # Or: head -c 48 /dev/urandom | base64 | tr -dc 'A-Za-z0-9'

Instructions for the implementer:

• First run the script to get a long random string of characters.
• Find inspiration from the string to determine direction: color palette, layout, typography, rhythm, metaphor. Look at sub-patterns, special numbers, letter shapes—anything that sparks ideas.
• Use your judgment to make that direction look good.
• The seed is only for inspiration; it must not appear in the final output.

For multiple options, each option gets its own new seed, generated in parallel, then choose one.

Technique 2: Bold, Specific Creative Brief

Replace "build a landing page" with requirements that are visual, for example:

• "Each section looks like a frame from a pixel-art game, yet remains a functional landing page."
• "Place it inside a living 3D isometric mini-city, where each feature is a block."
• "Aggressive asymmetry, jarring colors and fonts, unsettling white space—break all the rules but still look good."

Three-step method to find ideas:

1. Ask the model to list a wide but shallow range of design language directions, each in one sentence.
2. Pick a few you like, write your real reaction (what you like, dislike, where it's cliché), let the model tighten according to your taste.
3. Once satisfied, let the model write a concise prompt for the implementation agent to build the first prototype.

Principle: Feeding AI ideas directly back to AI only yields what everyone has; human selection is the source of uniqueness. Ideas that sound "this can't possibly work" are worth trying; failed prompts should be saved to retest when new models arrive.

---

II. Define

Technique 3: Independent Review Sub-Agent (Positive Feedback Loop)

The implementer is not objective—it can see its own code and reasoning. Open a separate "design review":

Each round:

1. Take a screenshot of the current design (desktop + mobile width).
2. Call the review agent in a completely new context, providing only the screenshots; no code, implementation details, previous versions, or past reviews.
3. The reviewer first describes what aesthetic the design is aiming for, then imagines what a top-tier design studio would do, and lists the biggest gaps.
4. Finally, give a score from 0-10: how close to that studio's quality.

The review prompt should specify:

• Evaluate both overall structure/composition and details.
• Deduct points for overused, flashy, or instantly recognizable AI-generated patterns.
• Feedback must be specific, concise, no empty talk.
• Be bold and opinionated, do not choose the safe or easy answer.

Rules:

• Stop conditions are only written for the implementer (e.g., "the review must independently score 9 or above to finish"), not in the review prompt, to keep the review objective. Use the same review prompt each round.
• Run 1-2 rounds first to see if converging, then decide whether to add more rounds, to prevent infinite token consumption.
• Prefer using a larger, more tasteful model for the reviewer; the implementer can use a cheaper, faster model but not so small that it cannot execute the direction. The reviewer only makes decisions, so its usage is small.
• The review criteria should be as objective as possible. Ideally: "Here are 5 images, 4 professional works + 1 ours, sort them by refinement and taste." Reference images can be used as a mood board, explicitly stating "use as benchmark, not to copy."

Technique 4: Enrich Visuals with Image Generation

Coding agents love using gradients, geometric shapes, and simple patterns as placeholders for images—this is exactly an AI trace. Actively demand:

• Use image generation for hero visuals, textures, illustrations, product images; can combine with shaders or 3D effects.
• Check actual results frame-by-frame in the browser.
• API keys are only for local development use, create a separate limited key, store it in a file that is git-ignored (e.g., .env.agents), and note in AGENTS.md "For development only, not to be shipped." Never hard-code it in code or final product.

Technique 5: Use Video Generation for Advanced Motion Effects

• Animated elements: Generate seamless loop clips on a plain background, then use chroma-key or a video matting model to remove the background and overlay onto any part of the interface, so it doesn't look like a video. For effects like glass refraction, first render onto the page's background color (baking in the refraction) before matting.
• State transitions: Generate transition clips by interpolating between start and end keyframes; the last frame of the previous segment becomes the first frame of the next segment, ensuring seamlessness; user scrolling or gesture can scrub frame-by-frame.
• Video models update quickly, can use an aggregation platform with one key and let the agent choose the appropriate model.

---

III. Deliver

Technique 6: Subtract

AI only adds, rarely deletes. Restraint itself looks premium. Ask for every element: "Does this really need to be on screen?" Common things to remove:

• Background glows, glowing progress bars, meaningless gradients.
• Arbitrary colored emphasis text and highlights.
• Redundant labels, containers, and white space when an image already expresses the idea.
• Custom buttons and input fields that are uglier than the platform's native components—prefer native components.
• Over-explaining caption text.

Deleting code is a "risk" for the model; you need to explicitly push it.

Technique 7: Remove AI Traces (This section is an original list from this repository)

Before delivery, check each item and fix if hit:

• Purple/blue-violet gradients, glassmorphism card stacks, large rounded corners and shadows everywhere.
• Universal hero area: left text + right image + top bar + centered main button.
• Three equal-width "feature cards" with linear icons.
• Emoji used as icons; meaningless decorative blobs and grid backgrounds.
• Generic title patterns like "Unlock your...", "Built for..." and empty marketing words.
• The same rhythm in every section: small label, big title, subtitle, button.
• Fonts only system sans-serif; type scale either too many or too flat.
• Faked statistics, testimonials, logo walls (either delete or replace with real data).
• Insufficient contrast, missing focus indicators, overflow on mobile—accessibility issues are also "unprofessional."

---

Execution Checklist for the Agent

1. Clarify the purpose, audience, and any brand or constraints that must be preserved (if none, don't force them).
2. During ideation: each option gets one random seed + one specific creative brief; generate 2-4 options in parallel, screenshot them for human selection.
3. After selection, enter the review loop: screenshot → new context review → modify → screenshot again; first 1-2 rounds, check convergence.
4. Add image/video assets when needed, verify frame by frame.
5. Subtract + check against the AI traces list + accessibility check (body text contrast ≥4.5:1).
6. Deliver: final screenshots, score trajectory from reviews, seeds and briefs used (for reproducibility).

Safety

• Do not write any API keys into the final product, repository, or logs.
• When generating assets, mind copyright: do not ask to imitate specific living artists' works or trademarks.
• Reference images are for benchmarking only, do not copy others' designs.