How to turn your AI into a world-class designer
Anshu Chimala · article · 2026-09-01 Reader 🖥️
Is AI Creative? (Pattern Recognition vs. Genuine Creativity)AI SlopMulti-Agent OrchestrationPattern-Completion Generation (Not Reasoning)Prompting as CommunicationTaste & Judgment
AI models default to bland, repetitive ('design-by-committee') design because their training rewards the statistically safest choice at every token, so producing distinctive AI-generated design requires deliberately counteracting that bias at each stage of a Discover/Define/Deliver process — injecting externally-sourced variety and ambition, judging output with a fresh-context critic instead of the generating agent, enriching it with image/video generation, and then subtracting rather than adding to reach a polished result.
practical · mixed — A technique listicle (Discover/Define/Deliver stages, seven numbered techniques) wrapped in a mechanism-style diagnosis (why RLHF-trained LLMs default to bland design) and licensed by the author's practitioner authority (12 years leading Apple AI/UX R&D). Every technique ships a verbatim, reusable prompt and before/after screenshots, which is more concrete than a typical listicle — but every demonstration is a single-author anecdote with no controlled comparison, blind evaluation, or replication, so treat the mechanism claims as plausible narrative rather than established fact. The captured content cuts off mid-way through Technique 7 ('Remove AI tells') behind a paid-subscriber paywall, so the piece is read as six complete techniques, not seven.
Structure
| Part | Role |
|---|---|
| Diagnosis: why LLMs make bad designers | premise |
| The Discover/Define/Deliver framework (adapted Double Diamond) | organizing structure |
| Discover — Technique 1: seed strings for genuine variety | procedure step |
| Discover — Technique 2: ambitious, specific prompts + AI-assisted idea refinement loop | procedure step |
| Define — Technique 3: design-critic subagent feedback loop | procedure step |
| Define — Technique 4: image generation to enrich designs | procedure step |
| Define — Technique 5: video generation for motion and transitions | procedure step |
| Deliver — Technique 6: cut elements that don't add value | procedure step |
| Deliver — Technique 7: remove AI tells (title only; content paywalled) | procedure step (incomplete) |
Key terms
- Discover / Define / Deliver
The author's three-stage design process for a team of AI agents, adapted from the human-centered Double Diamond model: explore broadly (Discover), give the design an individual identity (Define), then polish and productionize it (Deliver). - String Seed of Thought
Have the model generate a long random alphanumeric string via shell script, then interpret its subpatterns as unrevealed creative-direction input, so each design run is grounded in a genuinely external, non-repeating source of variation rather than the model's own (non-random) guess at 'randomness.' - design critic
A second AI agent instance, invoked in a fresh context with only a screenshot of the current design (no code, implementation history, or prior critique), that scores the design against explicit, objective criteria and whose 9-or-above verdict is the stopping condition for iteration. - AI tells
Visible markers that reveal a design was AI-generated — e.g. overexplaining, decorative elements with no practical purpose, code-native gradients/shapes instead of generated imagery — that Technique 7 promises to teach removal of. - design-by-committee
The author's metaphor for what RLHF-style, human-rated next-token prediction produces: a model that, at every design decision, fills in whatever is statistically most likely to please everyone, converging on bland, repetitive, safe output.
Propositions
- explicitLLMs are next-token predictors whose outputs are shaped by human ratings fed back into training, which teaches the model to make consistent, safe choices that fit everyone's preferences — so at each design decision it fills in the most statistically likely (safest) token, producing repetitive, bland, 'design-by-committee' results.conf: mediumextendsAre We Entering a New Age of Creativity with the Help of AI? · p6 — Supplies a specific training-mechanism explanation (RLHF-style human-rated next-token prediction converging on safe/average choices) for the homogenizing effect that the HBR/Atlantic-OpenAI doc asserts empirically ('AI's current effect on creative output is to homogenize and elevate less-creative people toward a higher floor').
- explicitSimply instructing a model to 'be unique' or 'make every design decision completely at random' does not produce genuine variety, because the model can only predict tokens that sound random, not act randomly.conf: high
- explicitSeeding the model with an externally generated random string (String Seed of Thought) as unrevealed creative-direction input produces genuinely varied, non-repeating design outputs, because the randomness is sourced from outside the model's own token predictions.conf: medium
- explicitHighly specific, ambitious prompts that supply a concrete external inspiration (a video game, an interior-design trend, an art installation) give the model a clearer creative target and produce more distinctive results than vague requests for uniqueness.conf: medium
- explicitAn agent asked to evaluate and improve its own design output is not objective, because it is anchored to its own code, past decisions, and prior rationale, and therefore cannot 'zoom out' the way a fresh reviewer can.conf: highsupportsAnthropic's Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development · p1 — Concrete design-domain confirmation of the general claim 'AI agents consistently overrate their own outputs, especially on subjective tasks like design' — this piece's Technique 3 is built entirely around that failure mode.
- explicitHaving a separate, freshly-initialized critic agent evaluate only a screenshot (not the code, implementation details, or prior iterations) against explicit criteria, iterated until it independently scores the design 9/10 or higher, produces designs with more individual identity than direct self-improvement loops.conf: mediumextendsAnthropic's Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development · p2 — Worked, design-specific application of the general principle 'separating the agent doing the work from the agent judging it is a strong lever for improving output quality' — adds a fresh-context-only, screenshot-scoped critic protocol with explicit stopping criteria.
- explicitConcrete, objective critic criteria (e.g. 'rank these 5 designs, 4 professional examples and 1 of ours, by polish and taste') produce more consistent evaluation than subjective criteria (e.g. 'judge if this looks beautiful, not AI-generated').conf: high
- explicitPairing a cheap/fast implementer model with an expensive/strong critic model is cost-effective, because the critic is invoked only for periodic 'executive' judgments rather than the bulk of generation; in the author's demos the critic (Fable 5) accounted for under 10% of output tokens.conf: low
- explicitCoding agents default to code-native visual effects (gradients, shapes, basic patterns) rather than incorporating generated images, and this default is itself a recognizable marker of AI-generated design.conf: medium
- explicitAI models default to adding more (over-explaining, decorative or non-functional elements) rather than removing, because stripping down a design and deleting code is a 'risky' choice the training-induced safety bias avoids by default — so most of the human effort in polishing AI-generated design goes into subtraction, not addition.conf: high
Arguments
To get distinctive (non-generic) AI-generated design, you must actively counteract the model's training-induced bias toward safe, average choices at every stage of the design process, rather than trust default prompting. strength: medium
Explicit premises:
- LLMs are trained via human-rated next-token prediction toward consistent, safe, broadly-pleasing choices (p1).
- Great design is the opposite of that — it bends rules and aims for an emotional response via unexpected choices (opening diagnosis).
- Simply asking a model to be random or unique fails to overcome this bias (p2).
Implicit premises:
- Design quality is at least partly recognizable/judgeable independent of which specific model produced it (i.e. a human or a critic model can validly assess it).
- A human director with practiced taste/judgment remains necessary throughout and is not itself replaceable by the generating model.
Inferences:
- If the bias originates in training-time averaging, then countermeasures must inject variance or judgment from outside the single generation pass — external random seeds, external ambition/inspiration, a separately-initialized critic, external image/video generation tools, and human-directed subtraction.
Design quality should be judged by a separate, freshly-initialized 'critic' agent instance rather than by the agent that produced the design. strength: medium
Explicit premises:
- An agent reviewing its own output is not objective, because it is anchored to its own code and rationale (p5).
- Objective, concrete criteria produce more consistent critic judgments than subjective ones (p7).
Implicit premises:
- A model without implementation context can perceive holistic/aesthetic quality better than one saturated in implementation detail — i.e. 'distance' improves judgment, analogous to a human 'fresh eyes' design review.
Inferences:
- Analogized directly from the author's stated experience managing human design teams ('zoom out, look at the big picture, think different') to a multi-agent LLM architecture.
Narratives
- Four instances of Claude Code given the identical bare prompt ('Build me a landing page for my productivity app') all return a purplish gradient, text-left/graphic-right layout — the default 'AI slop' pattern.
→ Establishes the baseline problem the rest of the article's techniques are built to escape. - The author's own calorie-tracking app: an initial Claude-generated design with pink glows, random highlight colors, redundant labels and custom controls worse than native iOS components, reduced via explicit subtraction instructions to an image-centric, Apple-native minimalist grid.
→ Concretely demonstrates the Technique 6 'subtraction' thesis with a real before/after from the author's own build, and implicitly licenses his authority as a former Apple design lead ('to my trained eye'). - The author's early career leading an Apple R&D team on exploratory future AI products, where 'preconceived notions... limited our creativity' until rigor and new process pushed the team to 'the fringes of what's possible.'
→ Transfers the author's human-design-team management credibility onto the AI-agent-orchestration framework that follows — it is the article's implicit warrant for treating human design-management lessons as portable to managing AI agents.
Questions addressed
- ✅ Why do some people's AI-generated designs look distinctive while others get generic 'slop' from the same models?
Training-time bias toward safe, broadly-pleasing choices, correctable via specific techniques that inject external variety/ambition, adversarial critique, and generative-media enrichment, then subtracted down. - ✅ How do you get a model to introduce genuine (not merely simulated) randomness into a design?
String Seed of Thought: have the model generate and privately interpret an externally-random alphanumeric string as creative direction. - ✅ How do you get an agent to objectively judge and improve its own design output?
Don't — use a separately-initialized critic agent shown only a screenshot, scored against explicit objective criteria, iterated to a fixed quality threshold. - ⭕ What exactly are 'AI tells' in a design, and how do you remove them?
Technique 7 promises to address this but its content is behind a paid-subscriber paywall in the captured version — genuinely unresolved for this reader. - ⭕ What is the real cost/iteration/time budget of stacking all seven techniques together on a realistic project (versus the headline 'built in three prompts' framing)?
Never addressed — the article never reconciles how many of the seven distinct techniques were actually stacked in its 'two or three prompts' hero demos. - ⭕ Do aggressively 'bold, rule-breaking' designs (the explicit goal of Techniques 1-2) trade off against usability/accessibility for real users?
Not addressed at all.
Gaps
- incompleteThe captured article cuts off mid-way through 'Technique 7: Remove AI tells' behind a paid-subscriber paywall ('This post is for paid subscribers'); the technique's content and any concluding wrap-up are unavailable to this reader.
Read next: Check Anshu Chimala's own Substack for a non-paywalled or later-unlocked version of the full piece before treating the technique list as complete. - uninformedEvery demonstrated result (seed-string variety, critic-loop convergence, the 10%-of-output-tokens cost claim, the calorie-tracker before/after) is a single-author, unreplicated anecdote with screenshots chosen by the author himself — there is no blind evaluation, no comparison group, and no measurement of how often these techniques fail or produce worse results than a baseline prompt.
Read next: Look for independent practitioner writeups or controlled comparisons of prompt-injected-randomness and critic-subagent design techniques — cross-check against this corpus's multi-agent-orchestration cluster (e.g. extract-anthropics-three-agent-harness) for whether the generator/evaluator pattern holds up outside a single anecdote. - incompleteThe article never reconciles its headline framing ('built in three prompts,' 'built in two prompts') with the seven distinct named techniques it goes on to describe — it's unclear how many techniques were actually stacked to produce the opening hero demos, or what the realistic cost/iteration/wall-clock budget of the full pipeline is for a team that isn't the author.
Read next: Treat 'N prompts' claims skeptically until technique count and iteration count are reconciled; look for a follow-up or reader Q&A addressing total cost. - illogicalThe central mechanism claim — RLHF-style training causes consensus-seeking causes generic design — is stated as settled fact via an intuitive 'design-by-committee' analogy, with no citation, benchmark, or reference to the actual mode-collapse literature it implicitly invokes.
Read next: Compare against more rigorous treatments of RLHF-induced output homogenization/mode collapse and the empirical AI-creativity debate elsewhere in this corpus (ai-creativity, pattern-completion-generation concepts).
Related documents
- Anthropic's Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development
- Are We Entering a New Age of Creativity with the Help of AI?
- AI Is Changing the Rules of Entrepreneurship
- The best response to AI slop, infinite advice, and online noise is from Robin Williams
- Harness design for long-running application development
- Charlie Hills 🦩's Post
- Presentation: Getting Rid of LeetCode Interviews in the World of AI
- No Vibes Allowed: Solving Hard Problems in Complex Codebases – Dex Horthy, HumanLayer
- How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar
- How to Actually Prompt Fable 5, Straight From the Team That Built It