Day Planner AI

A tool that plans your week across multiple interests without burning out — built as a learning project pairing each feature with an AI engineering concept.

backend/     FastAPI — all the AI logic lives here
frontend/    Next.js — the daily-use interface

cd backend
python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env

Fill in .env:

• ANTHROPIC_API_KEY — from https://console.anthropic.com
• TODOIST_API_TOKEN — Todoist → Settings → Integrations → Developer → API token
• GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET — create an OAuth 2.0 Client ID (type: Web application) at https://console.cloud.google.com/apis/credentials, after enabling the "Google Calendar API" for the project. Add http://localhost:8000/auth/google/callback as an authorized redirect URI.

Run it:

uvicorn main:app --reload --port 8000

Then, once, visit http://localhost:8000/auth/google/login in your browser to connect your calendar (one-time consent flow; token is cached in token.json).

Docs / try endpoints interactively: http://localhost:8000/docs

cd frontend
npm install
npm run dev

Visit http://localhost:3000. It talks to the backend at the URL set in .env.local (NEXT_PUBLIC_API_BASE, defaults to http://localhost:8000).

• POST /api/plan — takes your interests, weekly time budget, energy notes, and optional feedback from last week, returns a structured weekly plan (not free text) using Anthropic's tool-use forced structured output.
• GET /api/todoist/tasks — pulls your current Todoist tasks (not yet wired into the plan prompt — that's a good next step).
• GET /api/calendar/events — pulls upcoming calendar events (also not yet wired into the prompt).
• POST /api/plan/push-to-todoist — writes a generated plan back into Todoist as tasks with due dates.
• GET /api/plan/wallpaper — renders the most recently generated plan as a bird's-eye PNG (7-day grid, colored by interest area) sized for a phone lock screen. 404s until POST /api/plan has run at least once (the plan is cached to backend/data/latest_plan.json, gitignored). Meant to be polled by a device-side automation rather than the frontend — on Android, a Tasker profile ("HTTP Request" → save PNG → "Set Wallpaper", scoped to the lock screen) is the simplest way to keep it fresh without a native app.

• Structured output via forced tool-use (app/services/planner.py)
• System prompt design encoding actual reasoning (pacing, variety, burnout awareness) instead of a bare instruction
• Separating the LLM call, the data schema, and the API layer
• Real OAuth (Google) vs simple token auth (Todoist) — different auth patterns you'll hit constantly in real integrations

1. Wire Todoist tasks + Calendar events into the _build_user_message prompt in planner.py so plans respect what's already committed.
2. Add few-shot examples of "good" vs "bad" plans to the system prompt.
3. Store generated plans + user feedback somewhere (start with a flat JSON file, no need for a database yet) — this becomes your eval dataset.
4. Build a small eval script: given a plan + feedback, score it (rule-based first, then try an AI-as-judge prompt) for variety/pacing/burnout risk.