The sidebar was one of the most detail-dense sections to reconstruct. It contains multiple layered zones: a top logo row, navigation icons with labels, a recents list grouped by time (Today / Older), a user identity block at the base, and a collapsible state. All of this had to be mapped visually before any code was written.

The design process started with a hand-drawn layout on paper each section labeled, proportions estimated from screen observation, interactive states noted in the margins. From that sketch, every element was translated into Tailwind utility classes with careful attention to font weight contrast between section labels and list items, subtle background differences between active and inactive states, and the right amount of padding to keep the sidebar feeling open without wasting space.

The sidebar was built inspired by Anthropic's visual language — dark charcoal surfaces, near-white text on selected states, muted icon weights but every implementation decision was made independently without access to any official design tokens or specs.

• Logo Row

  Top-aligned Claude wordmark with sidebar collapse toggle icon.

• Navigation Icons

  New chat, Search, Chats, Projects, Artifacts, Code, Customize with label alignment.

• Recents List

  Chat history grouped into 'Today' and 'Older' with truncated titles and hover states.

• User Identity Block

  Avatar initials, display name, plan badge, and account actions pinned to sidebar base.

• Collapsible State

  Sidebar toggles to icon-only mode with smooth width transition.

• Active State Contrast

  Selected chat highlighted with subtle background lift and full-weight text.

The share chat feature was not just a UI detail it is a complete functional workflow. Users can choose between keeping a conversation private or generating a public link that anyone can view. This required building both the modal UI and the underlying sharing logic that generates, stores, and resolves share links.

The modal itself was rebuilt from observation: a clean dark overlay, two radio-style selection cards with lock and globe icons, descriptive subtext for each option, and a primary action button. The visual design mirrors Anthropic's characteristic low-contrast dark card pattern with a single high-contrast CTA — but every measurement was judged by eye from the actual interface.

On the backend, Supabase handles the persistence layer. Each shared conversation is stored with a unique slug, visibility status, and conversation snapshot. Public links resolve to a read-only view of the conversation. The sharing state is reflected in real time across the UI without a page reload.

• Private Mode

  Default state — conversation visible only to the authenticated user.

• Public Link Generation

  Generates a unique shareable URL stored in Supabase with a conversation snapshot.

• Read-Only Share View

  Public recipients see a clean read-only conversation page with no auth required.

• Real-Time State Sync

  Sharing status updates instantly across the UI without a page refresh.

• Modal UX

  Dark overlay, radio cards, descriptive labels, and a single high-contrast CTA — rebuilt from observation.

The artifacts panel is one of the most technically demanding features in the interface. When the AI generates code React components, HTML, SVG, Markdown the artifacts panel opens alongside the chat to show a live preview. The layout splits into a two-column view: the conversation on the left, the artifact panel on the right. The panel itself supports tabbed views switching between rendered output and raw code, with syntax highlighting and a copy action.

Building this required solving several challenges simultaneously. The split-pane layout had to remain fluid the chat column and artifact panel had to resize sensibly as the viewport changed. The code preview renderer had to safely sandbox and execute generated code inside an iframe without exposing the parent application. Streaming token output had to be parsed on the fly so artifacts appeared and updated in real time as the model responded, not after the full response was complete.

• Split-Pane Layout

  Chat and artifact panel rendered side by side with fluid column sizing.

• Live Preview

  Generated React, HTML, and SVG code previewed instantly in a sandboxed iframe.

• Code / Preview Toggle

  Tab interface to switch between rendered output and raw syntax-highlighted code.

• Streaming Artifact Render

  Artifact panel updates in real time as tokens stream in — no waiting for full response.

• Copy and Download Actions

  One-click copy and file download available directly from the artifact panel header.

• Markdown Rendering

  Markdown artifacts rendered as formatted documents with heading hierarchy and code blocks.

The settings page was one of the more structurally complex pages to rebuild. It uses a two-column layout with a persistent vertical navigation on the left General, Account, Privacy, Billing, Capabilities, Connectors, Claude Code and a scrollable content area on the right that loads the selected section. Getting the column heights and scroll behavior to work correctly across different viewport sizes without either column overflowing or clipping required careful handling of overflow and height constraints.

The General section is the most feature-rich. It includes an avatar uploader with initials fallback, a live editable display name, an email field, a work-type selector, and an open-ended custom instructions textarea where users can define how Claude should behave in every conversation tone, format, focus, language preferences. These instructions persist in Supabase and are injected into every new conversation as a system prompt prefix.

Beyond General, the settings architecture is designed to scale: additional sections like Billing, Capabilities, and Connectors have stub routes ready to receive real functionality as the product grows.

• Two-column layout — sticky left nav, scrollable right content
• Avatar uploader with initials fallback and image preview
• Live editable display name synced to Supabase user record
• Work-type selector for AI context personalization
• Custom instructions textarea — persistent system prompt injection
• Section-based routing: General, Account, Privacy, Billing, Capabilities
• Settings persisted in Supabase and applied per session
• Scalable architecture ready for Connectors and Claude Code sections

The pricing and plan upgrade page was designed and built in the same observation-driven way as the rest of the interface plan tiers, feature comparison, and a CTA structure all modeled from the actual Claude pricing layout. The page is currently live in the product as the gateway for future monetization. The plan system is still in active development.

The more technically interesting engineering behind the scenes is the AI model chain. This interface is not locked to a single API. Behind the model selector, a chained API fallback system is in place: Groq, Mistral, Cohere, Cloudflare Workers AI, Ollama, and OpenRouter models are all connected. When one API's rate limit is hit or a request fails, the next provider in the chain is automatically invoked — without any interruption visible to the user. The conversation continues seamlessly.

Prompt engineering is also built into the chain layer. Every conversation benefits from a structured system prompt that enforces consistent output formatting, response tone, and context awareness regardless of which model is serving the response. The result is a unified experience even when the underlying model changes mid-session.

• Plan Page

  Observation-rebuilt pricing layout with tier comparison and feature lists — future monetization ready.

• Groq Integration

  Primary fast-inference API for low-latency responses with high token throughput.

• Mistral & Cohere

  Secondary model tier automatic fallback when Groq rate limits are reached.

• Cloudflare Workers AI

  Edge-deployed model inference for geographic distribution and low-latency fallback.

• Ollama & OpenRouter

  Local and aggregated model access as tertiary fallback options in the chain.

• Prompt Engineering Layer

  System prompt injected at the chain level — consistent output formatting across all models.