Designing an LLM feature: this site's assistant

Role

Designer & builder

Timeline

Sep 2026

Platform

Web · Claude API

Domain

SaaS

The chat in the dock below: scoping, prompt design, tone, edge cases, fallbacks, designed and built by me.

Why a chat

A portfolio is static; visitors' questions aren't. A chat answers the question each visitor actually has: services, cases, fit. And where generative UI is NOT the answer, it stays out: navigation is classic, and the chat never blocks content.

Scope

Three jobs, nothing else: answer FAQ, talk about cases and experience, qualify leads toward email or a call. The system prompt is the design material: bio, cases, tone rules and behavior rules in one file I can edit and redeploy.

Guardrails

Rate limits per IP. Message length caps. Conversation memory lives only in the visitor's session, and nothing is stored server-side. If the API fails or the budget is hit, the assistant politely rests and hands over an email link. Cost budgeted to a few dollars a month.

What's next

Streaming answers, smarter follow-up suggestions from context, and honest analytics on what visitors actually ask.