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

Design cannot be one-shotted by AI; instead, tools should give users precise vocabulary (adjectives and verbs) to steer coding agents at the right level of c...

By Sean Weldon

Design at the Speed of Adjectives: Vocabulary as Control Surface in AI-Assisted Design Tools

Abstract

This synthesis examines the design philosophy underlying Impeccable, a portable design skill that augments AI coding harnesses including Claude Code, GitHub Copilot, Cursor, and Codex. The central thesis holds that design cannot be "one-shotted" by generative systems: it is a context-rich, multi-shot, multi-stakeholder activity that resists resolution through a single prompt. Rather than automating aesthetic judgment, the approach under study supplies users with an operational vocabulary - adjectives and verbs such as bolder, quieter, distill, and harden - whose meanings are explicitly defined in context and injected at identified points in a non-linear design workflow. Comparative demonstrations using identical prompts on the same underlying model indicate that semantically grounded vocabulary produces materially different outputs than unaugmented prompting. The work argues for amplified craft over automation, with direct implications for practitioners building human-in-the-loop creative tooling.

1. Introduction

The rapid adoption of agentic coding assistants has produced a visible convergence in generated interface aesthetics. Practitioners increasingly describe this phenomenon as AI slop - a recognizable set of default patterns signaling machine authorship. The pattern vocabulary itself evolves over model generations: numbered section layouts and purple gradients characterized earlier outputs, while more recent generations favor what is colloquially termed "Claude beige," paired with instrument serif typefaces and italic accents. The aggregate effect is termed algorithmic unilo: a homogenization of visual output across otherwise unrelated products, regardless of the underlying business or brand.

Impeccable is positioned as a direct response to this convergence. It is not a model, nor a standalone application, but a design skill - a portable layer of definitions and procedures that installs into existing coding harnesses and changes how those harnesses interpret design instructions. The governing claim is that the failure mode of AI-assisted design is not insufficient model capability but insufficient specification: users lack a vocabulary precise enough to steer agents at an appropriate level of control between manual manipulation and full delegation.

This synthesis addresses three questions: (i) why design resists one-shot generation; (ii) how vocabulary functions as a control surface between direct manipulation tools like Figma and fully agentic delegation; and (iii) why the tool's design explicitly refuses automation despite recurring user demand. Section 2 establishes theoretical context; Section 3 analyzes the core mechanisms of vocabulary-based steering; Section 4 extracts technical implementation details; and Section 5 situates the findings within broader industry trends around design-engineering role convergence.

2. Background and Related Work

2.1 The Collapse of the Handoff Model

The traditional waterfall progression from design artifact to engineering implementation is described as collapsing across organizations of all sizes. The design engineer - a practitioner operating fluently in both domains - is increasingly common. This role convergence creates a coordination problem: designers and engineers require a shared language that is simultaneously expressive enough to convey aesthetic intent and precise enough for machine execution. Impeccable is framed as supplying exactly this shared vocabulary, functioning identically regardless of which harness receives it.

2.2 The Lichtwort Concept

A key theoretical borrowing, attributed to Matt Puk, treats an adjective as a "light word" (German Lichtwort) - a term that is semantically hollow until deliberately infused with meaning. Applied to prompting, the implication is direct: "make it bolder" is not an instruction but a placeholder whose value depends entirely on whether the receiving system holds a defined, contextual interpretation of bolder. As stated in the source material, "an adjective with nothing behind it is just a nicer prompt." This concept forms the theoretical backbone for the tool's entire vocabulary system.

3. Core Analysis

3.1 The Two Extremes and the Missing Middle

The analysis identifies two poles of existing design tooling, both judged inadequate for contemporary work. At one extreme, direct manipulation tools such as Figma and Webflow operate at an elevation too low for most present-day tasks, requiring manual pixel-level intervention. At the other extreme, fully agentic delegation - issuing an instruction like "please build this" - produces generic, undifferentiated output. A viral tweet referenced in the talk, describing the difficulty of "centering a div with Opus," illustrates this tool-task mismatch: a frontier model applied to a trivial layout problem without appropriate scaffolding. Impeccable is positioned in the resulting gap, offering an intermediate elevation of control through defined vocabulary rather than raw prompting or manual manipulation.

3.2 Why One-Shot Generation Fails

The central empirical claim is that design is inherently non-one-shot. Even elite human design studios require iterative clarification - of emotional territory, target audience, and reference material - before production begins. A rapidly vibe-coded page can be technically competent while remaining aesthetically empty, because, as characterized in the source material, "nobody decided anything on this page." This diagnosis reframes AI slop not as a model quality problem but as a decision-absence problem: outputs default to statistical averages when no stakeholder has exercised judgment about hierarchy, tone, or constraint.

3.3 Vocabulary as Contextual Definition

The mechanism by which Impeccable addresses this gap is the explicit, contextual definition of design vocabulary. Commands including bolder, quieter, distill, polish, denser, and harden are not passed through to the model as bare adjectives; each carries a predefined operational meaning. For example, bolder is defined in terms of hierarchy, scale, and decisive typography - explicitly excluding gradients, glass effects, or neon accents, which are common AI slop defaults. The command harden is defined even more narrowly, addressing robustness and responsiveness across device classes rather than any aesthetic property. A demonstration using GPT-5.5 on an extra-high reasoning setting, with identical prompts issued with and without Impeccable installed, produced starkly divergent outputs, evidencing that the vocabulary layer - not the underlying model - accounts for the difference in design quality. A stated evaluative heuristic captures the intended threshold of quality: "Show someone your work and say AI made this bolder - if they believe you, you failed."

3.4 Workflow Mapping and Injection Points

Drawing on over two decades of tool-building experience, the design workflow is mapped as non-linear rather than sequential, comprising stages of initialization, shaping, crafting, iterating, hardening/polish, and return to the design system. Vocabulary commands are injected at specific points across these stages rather than applied uniformly. An additional command, overdrive, is designed to produce deliberately over-the-top output, demonstrated through a "radiant shaders" event-horizon shader project. This suggests the vocabulary set spans a range from restraint to excess, giving users graduated control rather than a single dial.

4. Technical Insights

Several implementation-relevant findings emerge from the analysis. First, the skill is harness-agnostic, functioning across Claude Code, GitHub Copilot, Cursor, and Codex, indicating that the intervention operates at the prompt/context layer rather than requiring model-specific fine-tuning or API access. Second, output divergence is attributable to context injection rather than model capability differences, since comparative tests held the model (GPT-5.5, extra-high setting) and prompt text constant. Third, the same harness produces substantially different results depending on whether the user's language signals designer versus non-designer intent, implying that the tool's value is partly diagnostic - inferring user expertise from word choice - and partly prescriptive, supplying missing definitional context. A key limitation is explicitly acknowledged: current AI is not capable of independently executing the final 5-20% of polish work separating good output from great output, though the tooling can accelerate the surrounding process. This bounds the tool's claims to augmentation rather than replacement.

5. Discussion

The refusal to implement a fully automatic mode is a deliberate design decision rather than an oversight. Requests for such a mode are described as frequent, and a pending pull request implementing automation is slated for closure. The stated rationale rests on the claim that taste can be amplified but not lab-grown: taste is contextual, cultural, and shaped by individual experience ("scars"), and any attempt to automate its production converges toward a muddy average once broadly replicated - precisely the algorithmic unilo the tool aims to counteract.

This positions Impeccable within a broader industry tension between efficiency-maximizing automation and judgment-preserving augmentation. As design and engineering roles continue to blur, tools that assume a shared, precisely defined vocabulary may generalize better across mixed-skill teams than tools assuming either pure manual control or full delegation. An open question for future investigation is how such vocabulary systems scale across cultural and organizational contexts, given that the meaning of terms like bolder is asserted to be context-dependent rather than universal.

6. Conclusion

The central contribution of this work is a reframing of AI design failure as a specification problem rather than a capability problem, addressed through explicitly defined, context-bound vocabulary injected at identified points in a non-linear design workflow. Practically, this suggests that tool builders seeking to support creative work with AI agents should prioritize precise, shared terminology over either raw manual control or unconstrained delegation, and should resist automating judgment even when users request it, since doing so risks eroding the differentiation that constitutes taste.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub