Vouch
ESLint for visual design — 45+ rules, 1ms per scan
The Problem
Your AI can generate a landing page in 30 seconds, but it cannot tell you the spacing is 13px instead of 16, or that your heading hierarchy skips a level. Teams shipping with AI copilots are building faster than ever — but visual quality drifts silently because nobody enforces design systems in production. There was no ESLint for visual design.
What It Does
- Scans live web pages against 45+ design rules across 14 categories: spacing, typography, color, contrast, touch targets, and more
- Surfaces exactly what has drifted and highlights where on the page
- Designed to plug into CI/CD pipelines or work as a click-and-see Chrome extension
What the AI Does
The linting is fully deterministic. 45+ hand-authored CSS rules running at roughly 1ms per scan. No model involved in the detection itself.
The AI layer sits on top as a command panel that executes context-aware tools: violation queries, fix suggestions, SARIF exports, and issue prioritization — with confirmation gates for any destructive action.
The AI is built but the extension is not yet released. The architecture is complete: streaming agent, tool execution pipeline, confirmation gates.
“AI can not tell you what good design is. But it can tell you how to fix the thing that is not.”
Key Design Decisions
Rules are human-authored, not AI-generated
What counts as a violation is an opinionated design call baked into code — not a probability score from a model. The AI interprets and explains; it never defines what good design is.
Design judgment is subjective and intentional. Letting a model decide what counts as a violation would remove the design opinion that makes the tool valuable.
Confirmation gates as workflow architecture
When the AI wants to mark something resolved, suppress a rule, or modify state, the user must confirm. The AI can suggest but cannot act unilaterally on any destructive action.
Autonomous resolution defeats the purpose. The designer needs to see and approve every change.
Deterministic-first architecture
The linter core is pure deterministic CSS analysis — zero API calls, zero latency. AI enhances results through the command panel but never gates the scan itself.
Vouch works offline, produces reproducible results, and can run in CI without an API key. Deterministic rules are auditable; black-box scoring is not.
What Happened
Built the linting engine, designed the command panel UI, completed the AI agent integration with streaming and tool execution, then moved to a new project before releasing. The Chrome Store submission is ready. The architecture is solid.
The product landing page at vouch.design was also designed and built as part of this project — a separate shipped artifact demonstrating the product positioning and visual identity.
Measuring Success
What to track when this resumes.
Suppression rate
If users suppress more than 30% of findings, the rules are too noisy and the tool erodes trust — the leading indicator of churn for any linting tool.
Violation-resolved rate
The percentage of flagged violations that users actually fix. Low resolution signals findings are not actionable.
Repeat-scan improvement
Does the same page score better on a second scan? Improvement over time is the clearest signal the tool is changing behavior.
Ready for more?
Prompt Theory
A Chrome extension for saving, organizing, and injecting prompts into ChatGPT, Claude, and Gemini with semantic search.
SG Resale: Agent-Native Retail
An AI-powered resale shopping experience — browse, discover, and purchase secondhand goods through an agent-native retail interface powered by Google Gemini.
Boon.town
An outcome-as-a-service platform that delivers submission-ready grant proposals assembled from your actual organizational documents.