Impeccable’s skill engineering aims to steer AI agents with designer vocabulary and human judgment
Impeccable’s Paul Bakaus argues that agents need domain-specific ‘skills,’ human-in-the-loop control, and a shared design vocabulary to avoid one-shot homogeneity and preserve creative ownership.
1 source · cross-referenced
- Impeccable introduces a design ‘skill’ system that translates designer terms like ‘bolder’ or ‘quieter’ into operational guidance for coding agents.
Paul Bakaus, creator of the open‑source design skills system Impeccable, argues that AI agents need more than instructions to produce high‑quality outcomes: they require domain knowledge, context, and ways for humans to steer the result. Impeccable gives coding agents a vocabulary to interpret designer intent—terms like “bolder,” “quieter,” “denser,” or “polished” are mapped to operational concepts such as hierarchy, scale, and typography rather than superficial effects like neon or glass surfaces.
Bakaus describes this as the emerging discipline of “skill engineering,” which treats skills as reusable components that encode professional judgment and constraints. He notes that most skills and models converge toward similar outputs when used generically, risking homogeneous designs; skill engineering aims to counteract that by embedding domain‑specific meaning into commands.
He also emphasizes that agent harnesses differ across models and tools (e.g., Codex vs. Claude Code vs. Cursor vs. GitHub Copilot), so a portable skill must account for varying capabilities and permissions. Bakaus has experimented with internal routing within skills—comparing it to a mixture‑of‑experts approach—to route tasks to the right sub‑skills or model capabilities while conserving tokens.
Impeccable’s live mode lets users select UI sections and request changes like “bolder” or “quieter,” with the system operating inside the project’s existing code and design system rather than exporting isolated mockups. Bakaus positions this as a potential “design harness” bridging chat‑based interaction and direct visual manipulation.
Bakaus rejects the idea of a fully automatic mode, advocating instead for agents to handle the first 80% of competent execution while humans steward the final 20% where taste, context, and distinctive point of view matter. He frames this as preserving ownership and purpose in creative work, countering visions that remove people from engineering altogether.
- Jul 3, 2026 · Latent Space — swyx
Debate at AI Engineer World’s Fair highlights tensions between agentic loops and engineering discipline
Trust76 - Jul 3, 2026 · Latent Space — swyx
Adobe demonstrates agentic websites that assemble pages in real time based on user intent
Trust78 - Jul 3, 2026 · Latent Space — swyx
Vercel’s Andrew Qu says agents are a new kind of software and unveils internal framework ‘eve’
Trust79