AI Engineer · Published 2026-09-10

Design at the Speed of Adjectives — Paul Bakaus, Renaissance Geek, Inc.

privateReview: unreviewedConfidence: high
design engineeringai workflowstypographytasteprototypingdesign-engineeringagent-controlvocabulary-interfacehomogenizationHuman-in-the-loopone-shot-limitsgenerative-craftai-slop
Open on YouTube ↗

Summary

Overview

  • Speaker: Paul Bakaus
  • Channel: AI Engineer
  • Main topic: AI-assisted design workflows, the limitations of one-shot design generation, and the importance of context, vocabulary, and human-in-the-loop control in design engineering.
  • Purpose: To demonstrate how AI can be effectively integrated into design workflows through vocabulary-driven controls (adjectives and verbs) rather than raw prompt generation or pixel-level manipulation. Paul Bakaus explores the intersection of AI and design, arguing against the illusion of one-shot design generation. He introduces Impeccable, a tool and methodology for designing at the speed of adjectives, emphasizing that AI requires rich context, iterative workflows, and appropriate human control points rather than total automation.

Topic Map

Introduction to Impeccable and the Problem with AI Handoffs

  • Explanation: Paul introduces Impeccable, a design skill for coding harnesses (Claude Code, GitHub Copilot, Cursor), and discusses the current failure modes of the traditional waterfall design handoff process in modern software development.
  • Key claims:
    • The traditional waterfall design-to-engineering handoff is broken and collapsing in small and large companies.
    • AI models frequently exhibit the 'AI slop' pattern, adding generic numbers, gradients, and neon without true design intent.
  • Examples:
    • A basic website generated via prompt vs. a website generated with Impeccable installed.
  • Terminology:
    • Impeccable
    • AI slop
    • Design engineer
  • Why it matters: Understanding why direct prompting without context fails to produce high-quality design outputs.

The Two Extremes of Design Control

  • Explanation: Analyzing the spectrum between direct pixel manipulation (Figma, Webflow) and full agentic prompting ('please design this for me').
  • Key claims:
    • Direct pixel manipulation feels too slow and outdated compared to agentic workflows.
    • Full agentic prompting ('one-shot design') results in generic, unopinionated outputs because nobody decided anything on the page.
  • Examples:
    • Centering a div with Claude Opus 4.7 as a meme example of mismatched tool usage.
  • Terminology:
    • Pixels
    • Prompts
    • Agentic workflows
  • Why it matters: Designers and engineers need to find the correct middle ground of control in AI-assisted creation.

Adjectives as Leitworte and the Workflow Map

  • Explanation: Introducing the concept of 'leitworte' (guiding words/adjectives) and mapping the non-linear design workflow from start, iteration, polish, to maintenance.
  • Key claims:
    • An adjective is a leitwort—a word infused with cultural, contextual, and design meaning.
    • You cannot one-shot design; effective design requires context-rich, multi-shot iterative workflows.
    • Taste cannot be lab-grown; it is cultural, contextual, scarce, and easily muddied by generic replication.
  • Examples:
    • Commands like 'bolder', 'quiet', 'distill', 'polish', 'denser', 'harden', 'overdrive'.
  • Terminology:
    • Leitwort
    • Context-rich multi-shot
    • Amplified craft
  • Why it matters: Establishes a vocabulary-based system for steering AI agents without breaking design systems.

Key Points

You cannot one-shot design

  • Explanation: Attempting to generate a fully polished, production-ready design with a single prompt results in empty, generic outputs ('AI slop').
  • Evidence: Comparison of single-prompt AI generation versus context-driven multi-step iteration.
  • Practical implication: Teams must embrace multi-shot, context-aware AI interactions that mirror human design processes.

Adjectives bridge the gap between design intent and AI execution

  • Explanation: Instead of manipulating pixels or writing raw code, using precise design adjectives (bolder, distiller, calmer) steers the AI agent correctly.
  • Evidence: Examples of shader adjustments and typography hierarchy scaling using specific design commands.
  • Practical implication: Design engineers should develop shared vocabularies that AI agents can parse and translate into design systems.

Taste cannot be automated or lab-grown

  • Explanation: Taste is cultural, scarce, and contextual. When everyone uses the same base models without unique input, everything converges into uniformity.
  • Evidence: The proliferation of generic purple gradients and identical modern SaaS aesthetics in early AI tools.
  • Practical implication: Human curation, brand identity, and deliberate constraints remain essential for distinguishing products.

Frameworks, Models & Processes

The Design Workflow Map

  • How it works: Mapping the non-linear design process into distinct stages with specific AI injection points.
  • Components:
    • Start / Initialization (scoping, shaping)
    • Iterate / Craft (adjectives, verbs, prototyping)
    • Polish / Harden (performance, responsiveness, visual polish)
    • Maintain / Design Debt Cleanup
  • When to use: When building or structuring AI-assisted design engineering tools and pipelines.

Examples & Case Studies

Comparing a baseline web page prompt in GPT-5.5 with and without Impeccable installed.

  • Illustrates: How structured context and vocabulary improve AI-generated layout, typography, and section hierarchy.
  • Lesson: AI requires guardrails and design systems to avoid defaulting to generic patterns.

Developing the 'overdrive' command in a shader library to push visual effects to the extreme.

  • Illustrates: How domain-specific verbs and adjectives can successfully push AI models to generate creative, non-standard outputs.
  • Lesson: Specific, community-tested commands yield much better results than vague prompt requests.

Actionable Takeaways

  • Immediate:
    • Stop expecting one-shot AI prompt generation to solve complex design problems.
    • Incorporate design vocabulary (adjectives and verbs) into your AI coding harnesses.
    • Map out your tool workflows before building or prompting.
  • Strategic:
    • Blur the lines between designer and engineer to create a unified design engineering workflow.
    • Focus on amplified craft rather than trying to fully automate taste and brand identity.
    • Build context-rich, multi-shot systems for AI interaction.
  • Questions to investigate:
    • What is the exact right level of human-in-the-loop control for AI design agents?
    • How can we prevent AI-generated interfaces from converging into a single homogenous aesthetic?
    • How do we scale design vocabulary across multi-person engineering teams?

Claims Worth Verifying

  • Taste is cultural, contextual, and scarce, and cannot be lab-grown. (philosophical claim)
  • One-shot design generation will remain unviable for high-end digital products due to lack of brand context. (predictive claim)

Notable Quotes

"Show someone your work and say AI made this bolder. If they believe you, you failed." "An adjective is a leitwort." "Taste can be amplified. It can't be lab-grown."

Compressed Summary

  • One-shot AI design generation produces generic outputs lacking genuine intent.
  • Impeccable bridges design and coding by using vocabulary-driven AI control.
  • Adjectives and verbs serve as guiding words (leitworte) to steer AI design agents.
  • Taste is cultural and contextual, requiring amplified human craft rather than lab-grown automation.
  • Keywords: design engineering, ai workflows, typography, taste, prototyping
  • Core insight: You cannot one-shot design with AI; true design engineering requires context-rich, multi-shot workflows guided by precise human vocabulary.

Core insights

6
Mechanismhigh noveltymoderate evidence

Design intent can be encoded as a constrained vocabulary of adjectives and verbs (leitworte) that the AI agent parses into design-system changes, forming a semantic control plane between human intent and code generation rather than pixel manipulation or open-ended prompting.

Why it matters

Gives a concrete alternative to both tool-level pixel editing and unbounded prompts: a small, shared, cognitively cheap vocabulary that steers an agent while remaining interpretable and reusable. This is a general pattern for any generative agent where the output space is too large for direct specification but too detailed for one prompt.

Generalization

Vocabulary-as-control-plane generalizes to any domain with a large, aesthetic- or policy-laden output space: audio mixing, motion, copy, API surface design. A bounded lexicon maps human intent to structured tool/DSL operations the agent executes.

An adjective is a leitwort—a word infused with cultural, contextual, and design meaning.
Open source video
Instead of manipulating pixels or writing raw code, using precise design adjectives (bolder, distiller, calmer) steers the AI agent correctly.
Open source video
Failure Modemedium noveltymoderate evidence

One-shot design generation is claimed to be structurally unviable for high-quality output because a single prompt forces no decisions to be made; the model defaults to generic 'AI slop' (generic numbers, gradients, neon) with no design intent.

Why it matters

Reframes the failure as a decision-allocation problem, not a model-capability problem: when the human delegates all decisions, the model has no constraints to resolve and converges on the statistical mode of its training distribution. Suggests agent design should force decision points rather than accept a single-shot spec.

Generalization

Applies to any generation task where quality depends on resolving many small trade-offs (architecture, copy, test suites). One-shot delegation yields mode-collapse to generic outputs; deliberate multi-turn constraint resolution raises quality.

Full agentic prompting ('one-shot design') results in generic, unopinionated outputs because nobody decided anything on the page.
Open source video
You cannot one-shot design; effective design requires context-rich, multi-shot iterative workflows.
Open source video
AI models frequently exhibit the 'AI slop' pattern, adding generic numbers, gradients, and neon without true design intent.
Open source video
Mental Modelmedium noveltyweak evidence

Taste is characterized as cultural, contextual, and scarce, and models sharing the same base distribution collapse toward a homogenous aesthetic (generic purple gradients, identical modern SaaS look) unless differentiated by unique human input and constraints.

Why it matters

Frames taste as a scarce, non-automatable resource and identifies homogenization as an emergent systemic risk of shared base models. Implies differentiation value accrues to whoever supplies unique context/constraints, and that 'sameness' is a measurable production hazard, not just an aesthetic complaint.

Generalization

Any generative system trained on a shared corpus exhibits convergence pressure; differentiation becomes a function of proprietary context, constraints, or post-training, not the base model. Relevant to eval design: detect mode collapse across outputs, not just per-sample quality.

Taste cannot be lab-grown; it is cultural, contextual, scarce, and easily muddied by generic replication.
Open source video
The proliferation of generic purple gradients and identical modern SaaS aesthetics in early AI tools.
Open source video
Taste can be amplified. It can't be lab-grown.
Open source video
Practicemedium noveltymoderate evidence

Reusable, named, community-tested commands (bolder, quiet, distill, polish, denser, harden, overdrive) act as a shared, extensible command library for agents; specific named ops beat vague prompt requests because they encapsulate tested transformation semantics.

Why it matters

Suggests design commands are better modeled as a versioned, testable operation library (a domain DSL) than as free-text prompts. Each op becomes an artifact that can be tuned, shared across teams, and improved collectively — a maintainable abstraction boundary for agent steering.

Generalization

Emergent pattern: agile vocabularies of named transformations become the interface to generative agents, the way shell verbs or refactoring commands abstract complex edits. Enables reuse, review, and regression testing of agent behavior.

Commands like 'bolder', 'quiet', 'distill', 'polish', 'denser', 'harden', 'overdrive'.
Open source video
Specific, community-tested commands yield much better results than vague prompt requests.
Open source video
Architecturemedium noveltymoderate evidence

Design work is modeled as a non-linear pipeline with distinct stages (Start/Init, Iterate/Craft, Polish/Harden, Maintain/Design Debt) that each expose different AI injection points, rather than a single monolithic generation step.

Why it matters

Provides an architectural boundary for where agent capability belongs across the lifecycle instead of a single 'design agent'. Different stages need different controls, context, and reliability guarantees, which affects harness design, memory scoping, and where human checkpoints sit.

Generalization

Stage-mapped injection applies to any agentic pipeline: scope the agent per phase (scoping, crafting, hardening, debt/maintenance) with phase-specific context and vocabularies rather than one general-purpose agent.

Mapping the non-linear design process into distinct stages with specific AI injection points.
Open source video
Start / Initialization (scoping, shaping)
Open source video
Iterate / Craft (adjectives, verbs, prototyping)
Open source video
Polish / Harden (performance, responsiveness, visual polish)
Open source video
Maintain / Design Debt Cleanup
Open source video
Practicemedium noveltyweak evidence

A proposed quality bar for AI-assisted output: if an observer, told the work was AI-made, still believes it was human-made, the craft succeeded; amplified craft (indistinguishable from human output) is the target, not visible AI speed.

Why it matters

Gives a falsifiable, human-eval-style acceptance criterion for generative quality — a Turing-style test for craft. Useful as a benchmark framing for evaluating whether agent output preserves brand or de-facto degrades to detectable AI patterns.

Generalization

Indistinguishability-from-human-craft as an eval criterion can be operationalized for any generative artifact (code, copy, UI): blind preference or detection studies across human/AI pairs.

Show someone your work and say AI made this bolder. If they believe you, you failed.
Open source video

Deep dives

5

Vocabulary-as-control-plane for generative design agents

Research question

Can a constrained, shared lexicon of design adjectives and verbs (leitworte) outperform open prompting and pixel editing in blind-rated craft quality, iteration speed, and design-system adherence?

Why

If validated, it establishes a versioned semantic interface between human intent and agent execution, generalizable beyond design to any large aesthetic or policy output space.

An adjective is a leitwort—a word infused with cultural, contextual, and design meaning.
Open source video
Commands like 'bolder', 'quiet', 'distill', 'polish', 'denser', 'harden', 'overdrive'.
Open source video
Source video

Decision allocation and the structural limits of one-shot generation

Research question

Does forcing explicit decision points and multi-shot iteration measurably reduce generic 'AI slop' outputs compared to one-shot delegation?

Why

Reframes one-shot failure as a decision-allocation problem, implying agent design should force constraints and choices rather than accept a single prompt.

Full agentic prompting ('one-shot design') results in generic, unopinionated outputs because nobody decided anything on the page.
Open source video
You cannot one-shot design; effective design requires context-rich, multi-shot iterative workflows.
Open source video
Source video

Homogenization and mode collapse in shared-base-model design

Research question

Can output diversity be measured across base models to confirm convergent aesthetics, and does unique context or constraint injection increase diversity?

Why

Treats homogenization as a measurable production hazard, affecting eval design, brand differentiation, and platform strategy.

Taste cannot be lab-grown; it is cultural, contextual, scarce, and easily muddied by generic replication.
Open source video
AI models frequently exhibit the 'AI slop' pattern, adding generic numbers, gradients, and neon without true design intent.
Open source video
Source video

Named design operations as a versioned agent DSL

Research question

Do named, community-tested commands with versioning and regression tests produce more consistent agent behavior than equivalent free-text prompts?

Why

Models design commands as a maintainable operation library rather than ephemeral prompts, enabling reuse, review, and regression testing.

Commands like 'bolder', 'quiet', 'distill', 'polish', 'denser', 'harden', 'overdrive'.
Open source video
Source video

Stage-scoped agent injection across the design lifecycle

Research question

Does phase-specific agent configuration (start, iterate, polish, maintain) outperform a monolithic design agent on quality and maintenance debt?

Why

Provides architectural boundaries for where agent capability belongs, affecting harness memory scoping, checkpoints, and vocabulary per stage.

mapping the non-linear design workflow from start, iteration, polish, to maintenance.
Open source video
Source video

Article ideas

4

The Adjective Is the Interface: Why Generative Design Needs a Versioned Vocabulary

Open prompts and pixel tools are both inadequate; the durable control plane for AI-assisted design is a constrained, versioned lexicon of adjectives and verbs that compiles to design-system operations.

Angle

Argue that the vocabulary layer—not the model—is the stable abstraction boundary teams should invest in, test, and share.

Source video

One-Shot Design Is a Decision-Allocation Failure

One-shot generation produces generic output not because models lack capability but because the prompt delegates every decision; quality rises when the workflow forces explicit decision points and multi-shot refinement.

Angle

Reframe 'AI slop' from a model-quality complaint into an agent-design pattern: build decision-forcing loops.

Source video

Taste Is Scarce: Homogenization as a Production Hazard

Shared base models exert convergence pressure toward generic aesthetics; differentiation must come from unique context, constraints, or post-training, and output diversity should be measured as a quality metric.

Angle

Treat identical SaaS aesthetics and purple gradients as measurable production debt, not just aesthetic taste.

Source video

Named Ops Beat Vague Prompts: Design Commands as a Versioned DSL

Community-tested named commands should be modeled as a versioned, testable operation library rather than free-text prompts, enabling regression testing and cross-team reuse of agent behavior.

Angle

Show how shell-like domain verbs become the interface to generative agents, with CI and semantic versioning.

Source video

Project ideas

4

Leitwort Control Plane Bench

new

A constrained, shared vocabulary of design adjectives and verbs compiled to deterministic design-system operations yields higher blind-rated craft quality and fewer iterations than open prompting or pixel editing.

Proof of concept

Build a coding-harness skill exposing commands like bolder, quieter, denser, harden that maps to design-token/component transformations; run on a set of UI recreation tasks against baselines.

Measurement

Blind expert preference, AI-vs-human detection rate, iteration count to acceptance, design-system adherence, output diversity.

Source video

Decision-Forcing Designer

movement-lab

Forcing explicit decisions for layout, type, color, and density across multi-shot generation reduces generic AI-slop patterns and increases perceived intentionality versus one-shot generation.

Proof of concept

An agent workflow that records and resolves decision points before and during generation; compare one-shot vs decision-forced outputs on identical briefs.

Measurement

Frequency of generic patterns (gradients, neon, generic numbers), human preference, decision trace completeness, revision count.

Source video

Homogenization Detector

beyond-evals

Outputs from shared base models exhibit measurable convergence that can be quantified; injecting unique context or constraints increases output diversity.

Proof of concept

Run identical design prompts across multiple models and seeds; cluster outputs by visual/style embeddings and detect generic-pattern signatures.

Measurement

Intra-model vs inter-model similarity, distinct style cluster count, frequency of known AI-slop patterns, diversity delta after context injection.

Source video

Named Ops Registry

gatehouse

A versioned registry of named design operations with regression tests yields more consistent agent behavior across contributors than free-text prompts.

Proof of concept

Implement a shared command registry (bolder, quiet, distill, polish, denser, harden, overdrive) with semantic versioning, test fixtures, and CI that checks token/component diffs on reference UIs.

Measurement

Cross-contributor output variance, regression pass rate, command reuse frequency, time to resolve command ambiguity.

Source video

Architectural implications

3

Control over generative output shifts from direct manipulation or open prompting to a constrained semantic vocabulary the agent compiles into design-system operations.

Before

Humans edit pixels in Figma/Webflow or issue open-ended prompts ('design this for me').

After

Humans emit leitworte/named ops (bolder, denser, harden) that the agent translates into structured design-system changes.

Consequence

The vocabulary layer becomes a versioned interface that can be shared, tested, and evolved independently of the model, creating a stable abstraction boundary between intent and execution.

Source video

The design-to-engineering handoff is described as collapsing, pushing design and engineering responsibilities into a unified design-engineering role.

Before

Waterfall design handoff: designers produce artifacts engineers implement.

After

A unified design-engineering workflow where the same actor steers agents across craft, polish, and maintenance.

Consequence

Tooling must support context-rich multi-shot interaction in the coding harness itself, not as a separate design tool; context and memory live alongside code.

Source video

AI injection points are distributed across pipeline stages rather than centralized in one generation step.

Before

A single design agent/generation call owns the whole artifact.

After

Phase-specific agent interventions (scope, craft, harden, maintain) each with their own context and controls.

Consequence

Harness design needs per-stage context scoping, distinct vocabularies, and stage-appropriate human checkpoints; a monolithic agent is the wrong abstraction.

Source video

Tradeoffs and failure modes

4

Automation vs taste/brand identity

Benefit

Agentic workflows let anyone generate production-grade layouts fast, collapsing the design handoff cost.

Cost or risk

Fully automating design converges outputs toward generic, brandless aesthetics because no unique constraints are applied; brand identity is lost.

Taste cannot be lab-grown; it is cultural, contextual, scarce, and easily muddied by generic replication.
Open source video
Source video

Pixel precision vs agentic speed

Benefit

Direct pixel manipulation offers exact control; agentic prompting offers speed and scale.

Cost or risk

Pixel tools feel too slow and outdated; open prompting yields unopinionated generic results — neither extreme yields both control and speed.

Direct pixel manipulation feels too slow and outdated compared to agentic workflows.
Open source video
Source video

Shared base models vs output differentiation

Benefit

Using common base models minimizes cost and effort.

Cost or risk

Shared base models cause output homogenization (identical modern SaaS aesthetics, purple gradients), eroding differentiation.

The proliferation of generic purple gradients and identical modern SaaS aesthetics in early AI tools.
Open source video
Source video

One-shot delegation vs multi-shot iteration

Benefit

One-shot prompting is fast and requires no iteration loop.

Cost or risk

Produces 'AI slop' because no decisions are made on the page; quality collapses without context-rich multi-shot interaction.

Full agentic prompting ('one-shot design') results in generic, unopinionated outputs because nobody decided anything on the page.
Open source video
Source video

Open questions

4

What is the exact right level of human-in-the-loop control for AI design agents?

Why unresolved

The summary frames the tradeoff between pixel-level manual control and full automation but does not identify or measure a specific optimal control point.

Research direction

Systematically vary human checkpoint density and vocabulary granularity across design stages; measure craft quality, iteration count, and homogenization.

Source video

How can AI-generated interfaces be prevented from converging into a single homogenous aesthetic?

Why unresolved

Homogenization is observed (purple gradients, identical SaaS aesthetics) but no mechanism to counter it is demonstrated.

Research direction

Test whether unique context injection, proprietary post-training, or constraint vocabularies measurably increase output diversity across models.

Source video

How can design vocabulary be scaled across multi-person engineering teams?

Why unresolved

Community-tested commands are reported to help, but governance, versioning, and consistency across teams are unaddressed.

Research direction

Prototype a shared command registry with versioning and regression tests; measure whether agents produce consistent results across contributors.

Source video

Does the indistinguishability-from-human-craft bar hold as a valid, measurable quality criterion for AI output?

Why unresolved

The criterion is stated as a rhetorical heuristic, not tested.

Research direction

Run blind human-detection/preference studies on AI vs human artifacts to calibrate the bar and detect AI-default patterns.

Source video

Key claims

7
causalVerification needed

One-shot AI design generation produces generic, low-intent outputs and is not viable for high-quality design.

Evidence

Full agentic prompting ('one-shot design') results in generic, unopinionated outputs because nobody decided anything on the page.

Question

Do controlled comparisons of one-shot vs context-rich multi-shot generation show consistent quality gaps under blind evaluation?

Source video
opinionVerification not requested

Taste is cultural, contextual, and scarce, and cannot be lab-grown or automated.

Evidence

Taste cannot be lab-grown; it is cultural, contextual, scarce, and easily muddied by generic replication.

Source video
comparativeVerification needed

Direct pixel manipulation and full agentic prompting are both inadequate extremes, requiring a middle ground of vocabulary-driven control.

Evidence

Direct pixel manipulation feels too slow and outdated compared to agentic workflows.

Question

Does a vocabulary-based control layer outperform both pixel editing and open prompting on speed and output quality in measurable tasks?

Source video
factualVerification needed

The traditional waterfall design-to-engineering handoff is broken and collapsing across companies of all sizes.

Evidence

The traditional waterfall design-to-engineering handoff is broken and collapsing in small and large companies.

Question

Is there survey or adoption data showing measurable decline in waterfall design handoff processes?

Source video
comparativeVerification needed

Named, community-tested commands outperform vague prompt requests for AI steering.

Evidence

Specific, community-tested commands yield much better results than vague prompt requests.

Question

Can controlled benchmarks show a statistically significant quality delta for named ops vs equivalent free-text prompts?

Source video
causalVerification needed

Shared base models drive convergent, homogenous output aesthetics.

Evidence

The proliferation of generic purple gradients and identical modern SaaS aesthetics in early AI tools.

Question

Can output diversity be measured across models to confirm mode collapse attributable to shared base distributions?

Source video
factualVerification needed

AI models default to adding generic numbers, gradients, and neon without true design intent ('AI slop').

Evidence

AI models frequently exhibit the 'AI slop' pattern, adding generic numbers, gradients, and neon without true design intent.

Question

Is the 'AI slop' pattern reproducible and measurable as a stable distributional signature across base models?

Source video

Connections

5