Set a duration or a deadline, then break the time into named activities. A live progress bar shows how deep into the session you are. When the session ends, get a full summary of how time was actually spent — and share it as a permanent public URL. An optional AI planner (BYOK) will suggest a session structure given a goal.
A fully client-rendered SPA hosted on Vercel. Session state lives in localStorage — no user accounts, no database. The AI planner calls OpenAI via a Next.js server action so the API key never touches the client. Completed sessions are serialised and encoded in a shareable URL.
This was my first project built with GitHub Copilot as a constant pair. The biggest surprise: Copilot was excellent at generating TypeScript interfaces and repetitive component scaffolding, but consistently suggested overly complex state management. I ended up simplifying most of its suggestions.
The shareable URL feature was almost entirely Copilot-written — it suggested the pattern, generated the encode/decode logic, and handled edge cases I hadn't considered. That felt like genuine leverage.
The AI session planner took longer than expected. Structured output from GPT-4o is reliable, but prompt engineering to get sensibly scoped suggestions took several iterations. The model tends toward optimistic time estimates.