Intel / post

Design Agent Harness

Source: x.com
Author: Neethan Wu
Date: Indexed Apr 17, 2026

Context

A reusable design setup can give an more concrete direction than a request to make a page look better. Engineer Neethan Wu describes a workflow that combines design instruction files, a visual canvas the agent can edit, and references Wu selects personally.

The saved instructions supply design conventions, while persistent specifications carry choices such as spacing and colors into later sessions. Wu uses the canvas to explore page designs and keeps the chosen direction as an implementation reference. Wu's role remains selecting and judging the result. This is an account of how Wu expanded a personal design workflow, with reusable guidance and visible iterations doing different jobs.

Key takeaways · AI-distilled

The tells that make AI-generated UI look AI-generated are specific and checkable: overused fonts, grey text on colored backgrounds, pure black, and cards nested inside cards. A design audit hunts exactly those.
Design canvases now come without a built-in model and drive your own agent over instead. One keeps the canvas as real HTML and CSS so there is no handoff step; another uses a Git-diffable JSON file that lives in your repo like code.
The cross-session amnesia problem — the agent forgetting every design decision you made last week — is solved by persisting spacing grids, palettes, depth strategy and component patterns to a file the agent loads automatically.
A style dropper extracts a design's visual DNA — palette, typographic rhythm, spatial density — and transfers it onto another layout, so browsing for inspiration ends in exportable code or a prompt you hand to your coding agent.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

Key quotes

“It catches the anti-patterns that make AI-generated UI look obviously AI-generated: overused fonts, gray-on-color text, pure blacks, nested cards.”

“The canvas is the shell; your agent is the kernel.”

“The first time I saw the agent swarm working visually on my canvas, it was mindblowing.”

“I spend about 20 minutes a day just scrolling through it. It's become part of how I warm up my eye before doing any design work.”

Read the source x.com