Skills, Guardrails, and the Craft Gap

It has been quiet here since June, and the P.S. explains part of why. I kept bookmarking through the gap, and three threads stood out: skills are replacing one-off prompts, design context is becoming the guardrail for coding agents, and the research is getting specific about where AI helps designers and where craft still falls short.

Skills Over Prompts

Framer Skills: Build better on-brand pages with skillsTurn the way you work into reusable skills for Framer’s agent. Define rules for design systems, content, SEO, and more, then apply them in any prompt.Framer

Framer Skills

Framer now lets you save how your team works as skills for its agent: design-system rules, SEO and launch checks, even a pass that strips out AI filler writing. You call them like commands, so /design-system builds a new page with the project's own type, colors, spacing and components. Write the brief once, reuse it everywhere.

Seamful Design Audit

A Figma Community skill that audits a design for the places it quietly decides for the user: defaults, compression, ranking, errors handled out of sight. Seamful design treats those seams as something to show, so people can see what the system did and adjust or push back. A useful check before shipping any AI feature.

GitHub - wheresryan22/anatomy: A Claude Code skill that turns a technical idea into a detailed, interactive isometric figure of a real machine. It makes SVG with optional WebGL shaders and 3D, for docs, articles and landing pages.A Claude Code skill that turns a technical idea into a detailed, interactive isometric figure of a real machine. It makes SVG with optional WebGL shaders and 3D, for docs, articles and landing page...GitHub

Anatomy, a Claude Code Skill

An open-source Claude Code skill that turns a one-line idea into an interactive isometric figure of a real machine, drawn in SVG with optional WebGL shaders and 3D. The examples range from a rocket engine to an arcade cabinet. If you write docs or landing pages, it is a fast way to get explainer diagrams without hand-drawing isometric art.

Context and Guardrails

Contextual AI Coding: How Agents Use Design Context | FigmaContextual AI coding means giving coding agents real design context to work from. Learn how Dev Mode, the Figma MCP server, and Code Connect make that happen.Figma

Contextual AI Coding

Figma's guide to giving coding agents real design context through Dev Mode, the MCP server and Code Connect. A written description of a button leaves the agent guessing; your tokens, component names and shipped code do not. The part I'd underline is that more context can hurt, so point the agent at what the task needs, not the whole library.

AI makes design system guardrails mandatory; this framework delivers them—Martian Chronicles, Evil Martians’ team blogAI has made implementing and enforcing design systems cheaper than ever. At the cost of speed, this falls through the cracks. But agentic development makes good design system practice non-negotiable. Implement a design system and bring consistency to an existing project.evilmartians.comAnton Lovchikov, Yuri Mandrikov and Travis Turner

Design System Guardrails for Agents

Evil Martians lay out an eight-point framework for keeping a design system consistent when agents write the code. Two ideas I'd take straight away: move deterministic checks out of AI skills and into scripts and lint rules, and give every component an explicit override prop that demands a written reason. Drift stops being silent and becomes something you can review.

What the Research Says

AI Tools Adoption across the Double Diamond Workflow: Phase, Mode, and Barriers in Designer PracticeDesigners are adopting AI faster than the tools built for them can keep up. This survey of 443 designers across 43 countries, among the first phase-disaggregated accounts of its kind, examined reported AI use across the four phases of the Double Diamond workflow (Discover, Define, Develop, Deliver). 79.7 percent reported confirmed AI use in at least one phase, but engagement was typically partial, spanning a mean of 2.89 of 4 phases, with adopters retaining the earliest phases and dropping the latest. Tool choice tracked each phase's dominant activity: conversational tools drove Discover and Define, AI-native image generation took over Develop, and Deliver showed a hybrid profile. AI-native and embedded AI use peaked in different phases, pointing to two distinct modes of human-AI collaboration: generating from scratch versus refining within existing software. Use intensity declined through fewer designers engaging, not scaled-back use. Adopters and non-adopters differed on one dimension: perceived usefulness.arXiv.org

Where Designers Actually Use AI

A survey of 443 designers in 43 countries maps where AI shows up across Discover, Define, Develop and Deliver. About 80% use it somewhere, mostly chat tools early and image generation in Develop, with use dropping off toward Deliver. Adopters and holdouts differed on only one thing: whether they found it useful.

Code2Video Benchmark: Evaluating AI Agents on Motion Design | HeyGen ResearchHeyGen Research releases the Code2Video Benchmark and a pairwise Judge model to measure how well LLM agents turn prompts into motion graphics code. Leaderboard, failure modes and dataset on Kaggle.HeyGen

Code2Video: Motion Design, Benchmarked

HeyGen Research tested leading models on turning 168 product-launch briefs into motion graphics code, scored on composition, timing and craft. Their summary: "Following the brief is nearly solved; craft is not." The common failures are the ones a motion designer catches first: hierarchy, scale, timing and pacing.

Beyond "ChatGPT Can Make Mistakes": Designing Interventions to Support Metacognitive Monitoring in AI-Assisted WorkAI assistance places a metacognitive demand on users, who must judge their own competence and the system's. Yet designers lack comparative evidence on which interventions to choose, where to place them, and how to tell whether they worked. We elicited 30 interventions from 11 experts and, with prior work, organized them into a design space of time (when an intervention acts), level (whose competence is judged), and source (who supplies the monitoring cue). A between-subjects experiment (N = 917; 12 planning-and-organizing problems) compared a per-task reliability card, contrasting replies, pause points, and post-problem reflection against a baseline LLM assistant. Reliability cards and contrasting replies reduced estimation error and overconfidence and increased aggregate confidence discrimination. No task-performance improvement or average within-item discrimination gain was established. We contribute a shared vocabulary, a design space, and evidence that measured monitoring and task performance are separable design targets.arXiv.org

Beyond "ChatGPT Can Make Mistakes"

A study of 917 people tested ways to help users judge both their own competence and the AI's. A reliability card per task and showing contrasting replies cut overconfidence, but neither improved task performance. If you design AI features, treat trust calibration and outcomes as separate targets, and measure both.

P.S. This newsletter moved off its old hosted platform to a setup I own. I built it with an AI coding agent over about two weeks and approved the risky steps myself. Every past issue, link and subscriber came over, and the switch took about a minute, so you likely didn't notice. It went smoother than I expected, mostly because the agent checked things instead of assuming. I'll share what it took and what I'd do differently in future issues. If you want that, hit reply, or subscribe at ai.ajmalafif.com.