Changelog

New Content

Prompt-template discipline: role / output / constraints / stop Ch 03
A reusable four-block prompt skeleton — role, output, constraints, stop — with five ready-to-run design templates (component spec, heuristic UX audit, task-flow/IA, token mapping, microcopy) and guidance for folding the skeleton into a skill you already run.

Visual prompting: point, draw, or talk on the live UI Ch 12
Cursor's Design Mode as an evolution of the review-and-iterate loop: annotate the rendered UI and the agent maps the mark back to source code, not a throwaway DOM patch. Links the official Design Mode announcement.

Generating ideas directly in HTML instead of walls of text Ch 04
HTML as a divergent-ideation surface — laying out the option space as a scannable page of variations rather than a prose list, with a worked prompt that forces breadth and layout.

html-video: a CapCut-style editing layer on Hyperframes Ch 09
An open-source editing layer on the Hyperframes runtime adding paginated preview, frame-level text editing, six auto-detected agent CLIs, and MiniMax-generated narration and music.

TypeUI: one markdown file per component Ch 03
A design-skill platform whose transferable lesson is the storage format: one markdown file per concern, served to multiple agents through one MCP endpoint, because agents read prose more fluently than schema blobs.

Multica: agents as assignable teammates on a task board Ch 11
A persistent-infrastructure orchestration model — agents as first-class task-board entries with squad routing, compounding skills, and cron autopilots — contrasted against the in-session orchestrator/subagent pattern.

Generative UI frameworks: when the agent renders the interface at runtime Ch 14
A future-direction note on AG-UI / runtime generative UI, weighing the reviewability cost: runtime-only interfaces trade diffability and versioning for immediacy.

Self-healing /goal commands and the limits of autonomous plans Ch 11
A community /goal upgrade that tries to detect when an autonomous run has gone off-plan and repair the plan, not just the code — flagged as a direction to watch and test.

Xiaohei illustration skill: a single-character explainer-graphic capability Ch 03
A skill that turns text into hand-drawn explainer graphics narrated by one recurring character, making the point that a skill freezes a complete visual fingerprint into a reusable instruction. Includes a figure generated live in the Xiaohei style for this edition.

SkillVault: a marketplace for discovering and distributing agent skills Ch 03
A cross-tool marketplace for skills, agents, hooks, and rules — framed as a discovery layer ahead of the install-and-review gate, not a substitute for reading a skill's source.

Infrastructure

• Bumped book version from 0.2.0 to 0.3.0; applied across chapters 03, 04, 09, 11, 12, and 14.
• Regenerated all output artifacts: self-contained HTML, paged HTML, PDF (320 pages), and EPUB.
• Generated a new Xiaohei-style explainer figure for the Chapter 03 illustration-skill section.

New Content

SkillSpector: Security Scanning for Agent Skills Ch 03, §03.5
NVIDIA's security scanner for agent skills checks 64 vulnerability patterns across 16 categories. Covers risk scoring (0-100 scale), installation, and integration with Claude Code, OpenCode, and Codex. Includes three real-world use cases: scanning community skills before install, auditing existing team skill libraries, and enforcing team security thresholds in CI.

Karpathy's CLAUDE.md: Four Rules, 65 Lines, 220k Stars Ch 03, §03.2.5
Andrej Karpathy's minimal CLAUDE.md hit number one on GitHub trending with 220,000 stars. This section analyzes why a 65-line file with four rules outperforms comprehensive but unfocused harnesses, and maps each rule to design-specific agent behavior.

Beautiful Feishu Whiteboard Ch 03, §03.4
A skill that generates editable SVG diagrams inside Feishu/Lark documents with 35 curated color palettes. A case study in platform-specific constraint handling: the agent must respect Feishu's renderer limitations and work around known export bugs.

anydesign: Extracting Design Tokens from Visual Sources Ch 04, §04.4a
A design skill that extracts structured design tokens from screenshots, live websites via MCP, and Figma files. Produces CSS custom properties, Tailwind config, and JSON token files with confidence levels for each extracted value.

From Reference Video to Animation Ch 09, §09.6
A new section covering the video-to-animation pipeline: how agents analyze reference video using ffmpeg frame extraction and LLM motion understanding, AnimSpec's video-to-prompt extraction service, MagicPath's agent-to-canvas handoff, and how all three fit into the Remotion/Hyperframes rendering stack.

frame.md: Design System Templates for Video Ch 09, §09.5a
A Hyperframes feature that bridges design systems and video output. Ships 8 templates, skeleton layouts, animation patterns, and shader transitions as markdown files.

Web Access: Firecrawl and Exa MCP Ch 10, §10.8
Two MCP servers that give agents live web access. Firecrawl provides structured web extraction at scale. Exa provides quick ad-hoc web searches inside agent sessions. Covers six practical workflows and security considerations.

McKinsey 24-Hour Sprint Framework Ch 14, §14.2
McKinsey's framework for agentic software delivery at organizational scale. Covers the daily cadence, eliminating handoffs through shared context, and building knowledge infrastructure.

Updated Content

Dynamic Workflows Ch 11, §11.6
Updated the trigger word from "workflow" to "ultracode" with explanation of the May 2026 change that eliminated false triggers. Added new subsection "Workflow Patterns from the Field" covering fan-out, adversarial verification, tournament, and loop-until-done patterns.

Infrastructure

• Updated publication date from May 2026 to June 2026 across all outputs.
• Bumped book version from 0.1.1 to 0.2.0.
• Generated new diagrams: SkillSpector pipeline, video-to-animation pipeline.
• Regenerated all output artifacts: HTML, paged HTML, PDF (317 pages), and EPUB.

Initial public release. 14 chapters covering agent platforms (Claude Code, Codex CLI, OpenCode, Gemini CLI), design-for-agents harnesses (CLAUDE.md, AGENTS.md, skill files), the design tool ecosystem (Claude Design, Open Design, OpenPencil, Huashu Design), design-as-code workflows, motion and video design (Remotion, Hyperframes), MCP integrations (Figma, Miro, MagicPattern, Paper), multi-agent design teams, and the future of autonomous design systems. 3 appendices with tool comparison matrices, MCP server references, and prompt templates. Approximately 47,400 words across 317 PDF pages.