AI UX DAILY
Wednesday, June 24, 2026
4 stories · curated for designers
The stories
Today in AI Products
 SKILL.md / Web Design Standard |
Jun 23 |
SKILL.md: Structured instruction sets that encode design judgment for AI agents
SKILL.md files are durable, versioned instruction sets that teach AI coding agents how to perform specific tasks according to professional standards, not just raw prompts. Unlike one-off prompts, skills encode judgment like preferring CSS Grid, using design tokens, and enforcing accessibility checks. This closes the gap between raw AI output and production-quality work by making standards shareable and repeatable across teams.
| “ |
Create a SKILL.md file in your design system or component library that documents your team's non-negotiable standards for AI-generated code (Grid over Flexbox, token usage, a11y rules). Version it like code and treat it as the source of truth for what "good" looks like to your agents. — Designer's Takeaway |
 Google DESIGN.md / Atlassian Testing |
Jun 22 |
Atlassian tested DESIGN.md against proprietary agent skills; token cost is real
Atlassian conducted production tests comparing Google's DESIGN.md format (a Markdown file for portable brand and UI context) against their own MCP server and agent skills. DESIGN.md consumed roughly 92% more tokens and produced greater output variance. The format works well for quick prototyping and customer theming, but falls short as a production standard compared to tighter, proprietary instruction methods.
| “ |
If you adopt DESIGN.md for AI context portability, budget for token overhead and plan fallback validation steps. Consider whether the interoperability benefit justifies higher costs, or whether a tighter, team-specific instruction format makes more sense for your shipping velocity. — Designer's Takeaway |
 Anthropic Claude Tag |
Jun 23 |
Claude Tag brings Claude into Slack with team-aware permissions and collaboration
Anthropic launched Claude Tag, a new Slack agent that lets teams collaborate with Claude directly inside Slack conversations. The agent respects team permissions and handles enterprise collaboration workflows. This moves Claude from a separate browser tab into ambient workspace context, embedding AI assistance into existing team communication and decision flows.
| “ |
Notice how Claude Tag integrates into existing team rhythm rather than forcing a new tool adoption. Consider where your design collaboration tools live (Figma, Slack, Linear) and whether embedding small AI helpers into those surfaces reduces friction more than shipping a standalone app. — Designer's Takeaway |
 Replit Agent |
Jun 23 |
Replit's evaluation framework: success means the finished app matches the vibe request, not test passes
Replit shared how they evaluate Replit Agent at scale. Most users start with nothing but a natural language idea, and Replit measures success by whether the finished app matches the request as a 'vibe coder' would judge it, not by whether unit tests pass or a patch applied cleanly. This reflects a fundamental shift in how AI output gets validated when humans care about the product experience, not the implementation details.
| “ |
When designing evaluation flows for AI-assisted work, separate technical correctness from user satisfaction. Build feedback loops that let people flag when outputs look right on screen but miss the intent, and use those signals to retrain your agent or adjust prompting strategy. — Designer's Takeaway |
Today's Idea
AI agents need design standards, not just prompts
The shift from "write me a button" prompts to durable instruction files (SKILL.md, DESIGN.md) reflects a maturation in how teams govern AI output. Token cost and variance matter, and tight instruction files outperform loose context files in production. Designers should think of these files as design system documentation that teaches AI what "good" looks like, versioning standards the way code gets versioned.
Stop shipping AI slop
Audit your AI design against 36 patterns
Drop a screenshot, get specific gaps and a Claude Code prompt to fix them. Free, no signup for the first audit.
Audit your design →