Field guide / primary domain
40
sources in this field
Updated July 16, 2026
Current thesis
The shortest path to orientation.
Brand design is being democratized through AI generation, but AI's weakness is structure not detail—it excels at filling UI details but fails at hierarchy, spacing, and layout. The fix constrains AI with professional foundations: layout templates, Dribbble references, dashboard patterns, and single-color-to-palette generators with contrast checking. Gemini produces full brand systems; OpenBrand extracts assets from URLs. Design taste becomes a retrieval-and-configuration problem rather than generation from scratch—reverse-engineer proven brands (Linear, Stripe, Vercel) into a design.md file that all modular skills reference. Google's open-source Design.md format captures brand DNA in markdown that AI agents query at generation time. Refero ships 2,000 searchable DESIGN.md files extracted from top products, solving why agents produce ugly UIs (they've never seen good design). High-quality UI remains a credibility multiplier making early-stage companies feel billion-dollar. shadcn/ui's adoption problem isn't components but theming—color palettes and font pairings matter more than individual components. Design systems now extend beyond UI to documentation artifacts (Vercel's Mermaid theme with animations). The AI Interaction Atlas emerges as a dedicated pattern library for human-AI interaction, signaling maturity of AI UX as its own design discipline distinct from traditional UI.
Evidence board
Claims worth carrying forward01Design.md is Google's open-source markdown format that encodes a brand's full design language (typography, colors, spacing) in one file. Attaching it to any agent prompt ensures every generated artifact—landing page, app, deck—inherits the same visual DNA without per-prompt re-specification.
02The recommended workflow: don't build a design system from scratch. Clone the design language of a brand you admire (Linear, Stripe, Vercel), extract it into a Design.md using ChatGPT or Claude, then layer specialized skill files on top (landing page skill, mobile skill, motion skill) that all reference the same Design.md.
03The critical failure mode in AI-assisted UI work: one screen looks polished, everything else looks generic. Design.md's structural fix is centralizing the aesthetic contract in one file so every subsequent generation—regardless of medium or agent—stays consistent.
04Build a second-brain capture habit for design inspiration: save visual references from the real world and web, then reference them when prompting new builds. Taste is developed through accumulated examples, not downloaded via a single prompt.
05River's core design constraint: it only works in public Slack channels and declines DMs. Every conversation is searchable company-wide. This single architectural decision—not the model—is what transforms an AI coding tool into an institutional learning system.
06River's PR merge rate improved from 36% to 77% over two months with no model change or retraining. The gain came entirely from employees watching River work, identifying failure points, and writing corrections into River's skills and memory. Collective human feedback beat model switching.
07River opened 1,870 PRs in one week in Shopify's main monorepo; ~1 in 8 merged PRs last week were River-authored. 5,938 employees worked with River across 4,450 Slack channels in 30 days. These are the scale numbers for a mature enterprise coding agent deployment.
08Lütke's thesis: the risk of AI isn't that it does the work—it's that it does the work in private windows and nobody else learns. When agent interactions are public, prompt patterns, investigation templates, and domain skills spread laterally without any training program or manager.
Adjacent fields
Latest evidence
GREG ISENBERG
Google's open-source Design.md format captures typography, colors, and spacing in a single markdown file that agents reference to produce consistent outputs. Combined with reusable skill files (landing page, mobile, pitch deck), it creates a design system any AI agent can apply uniformly across all surfaces—replacing the common pattern of polishing one screen while everything else looks generic.
tobi lutke
Tobi Lütke describes River, Shopify's internal AI coding agent living in Slack, designed with one hard constraint: it only works in public channels, not DMs. This forces all AI-assisted work into the open, creating a 'Lehrwerkstatt' (teaching workshop) where knowledge spreads osmotically. In 30 days, 5,938 employees used River across 4,450 channels; it opened 1,870 PRs in one week (~1 in 8 merged PRs). Merge rate climbed from 36% to 77% over two months purely from collective human feedback, not model changes.
elvis
A workflow pattern combining LLM Wikis (structured knowledge stores) with interactive HTML Artifacts creates a bidirectional agent-UI layer. Artifacts are built on top of wikis, can invoke agents, and agents can update artifacts—enabling inbox-zero automation, research scheduling, topic discovery, and live figure generation from a single HTML file.
Michael Guo
Michael Guo notes that Chrome Browser and Hyperframe are both available as Codex plugins, making them straightforward to incorporate into AI creative workflows.
Dave Kline
Dave Kline describes replacing a perpetually skipped manual Friday review with an AI-assisted workflow that completes in 12 minutes across 5 steps. The post teases the framework but the actual steps are not included in the captured content.
Peter Yang
A designer now working as a principal engineer has shifted to a fully code-native design workflow: AI generates a design spec (markdown), then components, then iterates based on taste. He considers this a core skill for any builder, not a replacement for Figma but a required complement.
OpenAI Developers
OpenAI's Codex gains a Record & Replay feature that lets users demonstrate a recurring workflow once—such as filing expense reports or submitting time-off requests—and have Codex convert that demonstration into an inspectable, editable skill. The user controls recording start/stop.
Nityesh
Hands-on Deck (EveryInc/hands-on-deck) is an open-source CLI tool (deck.py) that exposes PowerPoint manipulation to AI agents as atomic JSON patch operations rather than raw OOXML editing. The core insight is that giving agents a structured, linting-aware interface—inspect, apply, diff, render—makes them behave like designers (look, edit, verify) rather than programmers (write XML and hope). It installs as an Agent Skill in Claude Code, claude.ai, Codex, and any shell-capable agent.
Zephyr
Boris Cherny (Claude Code creator) no longer talks directly to Claude — he runs a system where Claude orchestrates other Claudes. This post outlines 7 features that close the gap between passive chatbox use and autonomous operation: Projects (persistent context), Connectors (live file access), Skills (reusable workflows), Artifacts (live tools), Memory (cross-session recall), CLAUDE.md (standing instructions), and Scheduled Tasks (timer-based autonomous runs).
Ole Lehmann
A workflow connecting Granola meeting transcripts to Claude Code via MCP to auto-generate context-aware dashboards post-call. The skill detects meeting type and produces tailored outputs: idea boards for brainstorms, timelines for planning, one-pagers for sales calls. Context is exported to an LLM wiki so Claude retains decisions for future sessions.