Impeccable: A Design Skill for Your AI Harness
Trending hard on GitHub today, pbakaus/impeccable, 32.6K stars, +612 in a day. It bills itself as 'the design language that makes your AI harness better at design.' Concretely: 7 domain reference files (typography, color, spatial, motion, interaction, responsive, UX writing), 23 commands (/impeccable audit, /impeccable polish, /impeccable critique), 27 deterministic anti-pattern rules, plus an LLM-based critique pass.
It works across Claude Code, Cursor, Gemini CLI, VS Code Copilot, basically any agent harness that supports skills/commands. The framing matters: AI-generated UIs all look the same not because models are bad, but because there's no shared design vocabulary loaded into the agent's context. Impeccable treats design as a learnable skill the agent can invoke, not just a vibe.
This is the same pattern as the COLLEAGUE.SKILL paper today and the Compound Engineering plugin trending last week. The harness layer is eating the AI stack. Taste, judgment, domain expertise all flow downhill into skill packages. If your product is 'AI does X but actually well', the answer is increasingly a skill, not a fine-tune.
The interesting validation here is the CLI tool that detects design problems without an AI harness at all. That means the 27 anti-pattern rules are explicit enough to be machine-checkable, which is also exactly what makes them teachable to an LLM. Deterministic guardrails plus LLM critique is the right composition: catch the obvious stuff with rules, use the model only where judgment is required.
Repo: https://github.com/pbakaus/impeccable
← Back to all articles
It works across Claude Code, Cursor, Gemini CLI, VS Code Copilot, basically any agent harness that supports skills/commands. The framing matters: AI-generated UIs all look the same not because models are bad, but because there's no shared design vocabulary loaded into the agent's context. Impeccable treats design as a learnable skill the agent can invoke, not just a vibe.
This is the same pattern as the COLLEAGUE.SKILL paper today and the Compound Engineering plugin trending last week. The harness layer is eating the AI stack. Taste, judgment, domain expertise all flow downhill into skill packages. If your product is 'AI does X but actually well', the answer is increasingly a skill, not a fine-tune.
The interesting validation here is the CLI tool that detects design problems without an AI harness at all. That means the 27 anti-pattern rules are explicit enough to be machine-checkable, which is also exactly what makes them teachable to an LLM. Deterministic guardrails plus LLM critique is the right composition: catch the obvious stuff with rules, use the model only where judgment is required.
Repo: https://github.com/pbakaus/impeccable
Comments