6 Agentic Patterns from Claude Design for Vertical Apps
Claude Design's edge comes from stacking 6 patterns—context grounding, structured memory, iterative multimodal refinement, self-QA, multi-variation generation, handoff—around a strong LLM like Opus 4.7. Build your legal, sales, or medical agents the same way: ground in user data first, then iterate with quality checks.
Ground Context Dynamically Before Generating
Never generate blindly—always ground your agent in user-specific data via progressive disclosure RAG. Claude Design forces users to build a detailed design system upfront (brand context, colors, fonts, HTML for buttons/cards), then the agent selectively pulls relevant chunks into its window per task, avoiding context bloat. Extend this: for sales agents, RAG over CRM data before outreach; for legal, load templated contracts before drafting. First output isn't the deliverable—it's a structured memory artifact (markdown, HTML/CSS, or JSON) that persists across sessions and downstream agents. This memory boosts speed/quality on repeats since models now parse these formats natively. Trade-off: initial setup time, but it crushes static system prompts, which most enterprise agents still misuse.
Iterate with Multimodal UX and Self-Critique
Ditch chatbot-only interfaces; let the agent generate task-specific controls (sliders for email aggressiveness, layout tweaks) from its own tokens, rendered dynamically. Claude Design handles 5+ inputs—chat, voice, DOM hovers, sketches, screenshots—for natural refinement loops. After output, run a self-QA: render (e.g., screenshot), feed back to vision model (needs strong one like Opus 4.7), critique/iterate until it matches intent. Burns tokens but yields production quality before user sees it—apply to emails, contracts, PDFs, UIs. This fixes where most agents fail: rigid UX and unpolished drafts.
Generate Variations Proactively and Handoff Seamlessly
Surface decisions hierarchically with multi-variation: output 2-3 options upfront (e.g., layouts > typography > colors), letting users pick axes like sales tone (warm vs. direct). Beats clarifying questions by showing trade-offs concretely, reducing back-and-forth. Finalize with handoff: export to portable formats (HTML/CSS, PDF, PowerPoint) for tools like Figma, Canva, Claude Code, or other agents. Avoid proprietary traps—markdown/JSON works everywhere. In multi-agent futures, this enables ecosystems; today, it matches user workflows. Stack all patterns (especially grounding + memory) for qualitative leaps—most deployments lack this dynamic harness.