Uncommon Waters
Academy
← 101 BuildWeek 4 · Nov 30 – Dec 4

Secrets, an AI feature, email, and watching production

env, LLM API, Resend, Sentry

Every key lives in env with a provider spend cap and none reach the browser; one key rotated live; one server-side LLM API call (system prompt + structured output) drafts on user data and waits for human approval before Resend sends the email; Sentry catches a deliberate error and the student gets the alert.

Monday · 90 minutes live

  1. 15′Frontier failure: eleven credentials pasted into chats in the founder record. A transcript is forever. Then a key found in the browser's network tab.
  2. 20′Secrets and cost: .env.local, Vercel env, server vs browser, provider spend cap, what rotation means. Calling an LLM API from the server: system prompt, structured output, human approval before effect.
  3. 35′Hands on: set a provider spend cap; wire one server-side LLM call that drafts from user data; human approves -> Resend sends the real email. Install Sentry, throw a deliberate error, watch the alert arrive.
  4. 15′Rotation drill: rotate a key live while the buddy watches; the app still works. Buddy opens the network tab and hunts for keys.
  5. 5′Habit check: keys never in chat. Today you proved why.

Tuesday to Thursday · you build

  • All keys in env, local and Vercel. Provider spend cap set. Secret scan clean.
  • One server-side LLM API call (system prompt + structured output) with a human approval step that sends a real email via Resend.
  • Sentry: one deliberate error caught, alert received, fixed in a PR that references it.

Office hours · two 30-minute blocks

  1. 1Tuesday: Resend domains and sandbox, env vars that do not load.
  2. 2Thursday: secret scan hits, rotate together; Sentry noise.

Exam whisper · five minutes of vocabulary

Anthropic Claude Academy: AI Fluency, first half. Exam Whisper: tokens and cost, system prompt vs. user prompt, structured output, agent = model + tools + loop, human-in-the-loop.