IGRIS · Production Website Agent + Quality Bar

AI DevelopmentAI Agents

The named craft intelligence behind every site in this portfolio. 9-step pipeline, 4-question quality framework, full tool registry across every modern framework. Mission: beat Awwwards, not match them.

IGRIS · Production Website Agent + Quality Bar
Anthropic ClaudeNext.js / Vite / AstroDifferentiation EngineAward-gradeProprietary

Overview

IGRIS is the craft-layer agent at the center of my build process. It's not a workflow doctrine — it's the named intelligence that picks the stack, defines the signature moment, enforces the quality bar, and ships. Every project starts with four mandatory questions: What's the ONE visual idea that makes this unforgettable? What would someone screenshot and post? What does the visitor feel in three seconds? What design decision here has never been done for this industry before? If I can't answer all four, the design isn't ready and we go back to research. The 9-step pipeline runs Brief → Research (Awwwards + Dribbble + competitor crawl) → Signature moment → Stack selection → Design system → Hero first → Full build → Preview audit → Deploy. The tool registry is deliberately broad — Next.js, Vite, Astro, SvelteKit, Remix on the framework side; GSAP, Framer Motion, Rive, Lottie, Anime.js, Three.js / R3F, Spline, Babylon, WebGPU on the motion + 3D side; Supabase, Neon, PlanetScale, Convex, Clerk, Sanity, Stripe, Mux, Cloudinary, Remotion, Mapbox, Expo on the platform side — because the stack must serve the design vision, not the other way around. Phase 1 (live now): world-class front-end across any stack. Phase 2 (in progress): full-stack autonomy — backend, database, auth, business logic, API layer all built and deployed automatically. Phase 3 (the goal): IGRIS spawns asynchronous autonomous sub-agents, each owning a domain (design, frontend, backend, database, auth, deployment), executing in parallel. A complete production-quality full-stack application in under five minutes. Source is proprietary — happy to walk through in a call.

Features

  • 9-step pipeline · Brief → Research → Signature moment → Stack → Design system → Hero → Build → Audit → Deploy
  • 4-question Quality Framework — mandatory before any code is written
  • Differentiation Engine — fingerprint registry refuses aesthetic repetition across builds
  • Full tool registry — picks per project, not per template (Next/Vite/Astro/SvelteKit; GSAP/Framer/Rive/R3F; Supabase/Convex/Clerk/Stripe)
  • Award-tier quality bar — every output benchmarked against Awwwards winners + Dribbble top shots before ship
  • Phase 2 active: full-stack autonomy (DB, auth, API, business logic)
  • Phase 3 vision: parallel multi-agent swarm, complete app in under 5 minutes

Technologies Used

  • Anthropic Claude (Sonnet / Opus) — agent intelligence core
  • Claude Code SDK + custom orchestration shell
  • Stack picked per project: Next.js / Vite / Astro / SvelteKit / Remix
  • Motion arsenal: GSAP, Framer Motion, Rive, Lottie, Anime.js, Three.js / R3F
  • Backend arsenal: Supabase, Neon, PlanetScale, Convex, Drizzle, tRPC
  • Custom differentiation engine + build_registry.json fingerprint store

Challenges & Solutions

Challenges

  • Defining aesthetic distance between two builds in a way the agent can enforce
  • Preventing drift into a house style under sequential-output pressure
  • Holding one award-tier quality bar across rapid output, not just one hero build

Solutions

  • Fingerprint extractor hashes palette + typography + motion grammar into a comparable vector
  • Differentiation gate at build-start refuses any spec matching the last N fingerprints
  • 4-question quality framework halts the build if any answer is unclear — back to research, never forward to code