planmyday

Overview

planmyday is a conversational daily planning tool that helps structure your day through dialogue. Unlike static planning apps with forms to fill out, it asks clarifying questions, pushes back on unrealistic schedules, and remembers your preferences over time.

Built entirely from scratch—no LangChain, no agent frameworks. Just Python, the OpenAI API, and a state machine.

What I Built

State Machine Architecture

The core of planmyday is a state machine that controls the conversation flow:

• •Idle — Waiting for user to describe their day
• •Questions — Asking clarifying questions about priorities and constraints
• •Planning — Generating a structured schedule
• •Feedback — Iterating based on user input
• •Done — Final plan delivered

This pattern makes the agent feel like a collaborator rather than a text box—it decides what to do next based on context, not just the last message.

Structured Outputs

Used OpenAI's structured output feature with Pydantic to guarantee reliable, validated responses. No regex parsing, no hoping the LLM returns valid JSON. The model either returns a properly typed object or fails cleanly.

Profile & Persistence

The agent learns and remembers:

• •User preferences and recurring commitments
• •Session context that persists between conversations
• •Historical patterns to improve suggestions over time

Check-in System

Added follow-up functionality to ask whether plans were actually executed—closing the loop between planning and reflection.

Why No Frameworks

Frameworks like LangChain are powerful but hide what's actually happening. When something breaks, you're debugging abstractions instead of understanding fundamentals.

Building from scratch meant understanding exactly how agents work: the loop, the state, the decisions, the memory. That knowledge transfers to any framework or tool.

Learnings

For a deeper dive into the design decisions and lessons learned building planmyday, read Building an AI Agent from Scratch.