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

Paul Bakaus, founder of Renaissance Geek, opens with an intentionally bland website and one small request: make its workflow section bolder. Impeccable is a design skill that runs inside coding harnesses including Claude Code, GitHub Copilot, Cursor, and Codex. With the skill installed, the section gains a more forceful visual hierarchy. The result still carries an unwanted generated-design tell—decorative section numbers—so the claim is improvement, not perfection.

The comparison then repeats the same prompt on the same project without Impeccable. The visible difference motivates the rest of the talk: a vague adjective leaves a general model to choose its own interpretation, while a loaded design command constrains what the adjective should mean. This is a qualitative demonstration judged from two outputs; the recording supplies no measured evaluation establishing the size or consistency of the improvement.

The old product handoff is already collapsing. Designers increasingly work in code, engineers make design decisions, and design engineers span both. That weakens the familiar waterfall in which product defines the work, design produces an artifact, critique follows, and engineering eventually implements it. Impeccable’s shared vocabulary matters in this environment because design intent must survive continuous cross-functional iteration rather than one formal handoff.

Current tools often pull people toward one of two extremes. Direct manipulation in Figma, Webflow, or another visual editor gives precise control over margins, padding, and the final appearance, but keeps the user at a low level of abstraction. Fully agentic delegation raises the level too far: “design this for me” leaves the agent responsible for choices the user never made. Using a powerful model merely to center a diff is comically excessive, yet handing it the whole design problem creates a different failure. The missing layer is steering at the level of intent.

Generated-design clichés also move. Purple gradients once made AI-produced pages easy to identify. Newer outputs may instead converge on beige palettes, serif display type, and italics—what Bakaus calls “Claude Beige” and “Algorithmic Uniqlo.” Any single page may be competent. The problem is sameness: when a model repeatedly draws from its median design reference, the output contains style without a project-specific reason for that style.

Effective design is context-rich and iterative. It depends on what the product is trying to accomplish, who will use it, and how several interested people respond to successive versions. Even a solo builder must inspect and revise the result; within a team, users and stakeholders add more opinions. A single agent call cannot wish that process away.

The first useful questions define the design territory: What emotional response should the product create? What must it never feel like? Which references are relevant? Who is the audience? A human design director could not responsibly accept “design my brand” and walk away without these answers. An agent needs the same missing context, even if it can produce a plausible page without asking for it.

This explains why a polished one-shot page can still feel empty. The model has rendered an image, but nobody chose what the hierarchy should communicate, what emotion belongs to the product, or why this visual language fits the audience. Competence is not the same as intention.

Impeccable places control in adjectives and verbs used during real design work. Each command names a direction rather than a pixel operation: bolder raises emphasis, quieter reduces it, distill simplifies, polish prepares the work for finishing, denser changes information density, and harden checks whether the design works across devices and conditions. Hardening may expose weak responsiveness or poor performance rather than merely changing appearance.

Designers and engineers can use the same model yet receive different results because their language carries different distinctions. Impeccable compresses some of that design vocabulary into a reusable skill, giving a less design-fluent user a way to express a direction the agent can translate into concrete work. The command does not replace the user’s decision; it makes the decision legible to the agent.

The “bolder” example shows the causal chain. Without added guidance, the model can interpret boldness as new colors, gradients, glass effects, or neon. Impeccable instead defines it through hierarchy, scale, and decisive typography. The observable change should be stronger attention and clearer emphasis without breaking the existing design system. The adjective remains easy to say, but its implementation is no longer unconstrained.

The command also carries a quality check: show the result to someone, say AI made it bolder, and “if they believe you, you failed.” The memorable phrasing pushes the agent away from conspicuous AI decoration and toward changes that look like a deliberate extension of the project. In practice, this asks the model to critique whether its own additions have become generic tells.

Borrowing a framing he credits to Matt Pocock, Bakaus describes an adjective as a light word that can be infused with domain meaning. A model already has many associations for “bolder”; the skill narrows those associations for this design task and project. That level of abstraction still has limits. It can help between the first exploration and the finishing pass, but Bakaus does not expect AI to replace human work in the final five, ten, or perhaps twenty percent needed to move from good to great.

A useful tool starts with the target user’s actual workflow. Design does not move through one clean line, so Impeccable identifies intervention points across shaping, crafting, iterating, hardening, and polishing. The work eventually returns to the design system, where reusable decisions are extracted and technical or design debt is cleaned up. Commands belong at different stages rather than being treated as interchangeable prompt decorations.

The stages support distinct jobs: shaping determines what should be built; crafting and iteration use directional words to explore and revise; hardening and polishing address edge cases and finish; design-system cleanup folds useful choices back into shared components and removes debt. Mapping these jobs prevents one command from pretending to solve the whole process.

overdrive is the playful exception that proves the mechanism. The command deliberately pushes a design over the top. Bakaus used it on an already developed event-horizon shader from Radiant Shaders to make the effect even more extravagant; it did not generate the shader in one shot. Community enthusiasm persuaded him to keep a command he had initially treated as half a joke. This is command design by observed usefulness and joy, not by completing an abstract taxonomy.

The caveat is simple: an adjective with nothing behind it is only a nicer prompt. Because the model has encountered many meanings for the word, the skill must define which meaning applies, which changes fit the project, and which familiar shortcuts to avoid.

The most requested feature is also the one Impeccable will not add: automatic operation. An auto mode that chooses every command and completes the design alone would recreate the one-shot workflow the tool was built to avoid. Bakaus is explicit enough to close an open pull request proposing it: “There is no auto, and there will be no auto.”

The refusal protects a product decision, not a missing implementation. Impeccable exists to let people steer toward an intended result. Removing the need to choose a command would also remove the moment when a person decides whether the design should become bolder, quieter, simpler, denser, or more resilient.

That leads to “amplified craft.” Tools can sharpen and extend a person’s taste, but Bakaus does not believe taste can be manufactured in a lab. Taste is contextual, cultural, and scarce; once everyone reproduces the same visual choices, those choices stop functioning as taste and become another muddy default. Impeccable therefore aims to increase the user’s ability to express judgment, not to solve judgment on the user’s behalf.