Three Notebooks

Inspiration

A task labeled “today” may actually belong to three different timelines:

• something left unfinished in the past,
• something needed to live or enjoy today,
• or something done today to create a future.

When unfinished work from the past fills the present, people lose the space to enjoy today or prepare for tomorrow. Three Notebooks was designed to return that time, attention, and emotional space to the user.

What it does

The user speaks or types a mixture of tasks, plans, feelings, and dreams. GPT-5.6 turns that input into editable candidates, then tentatively sorts each item by its cause and direction:

• Past: work that would not exist now if it had been completed earlier
• Today: daily life, recovery, rest, and enjoyment
• Future: actions that create something that does not exist yet

The user reviews and corrects everything before saving. Saved items update the ToDo list, timeline chart, and calendar. A White Note offers a space to think with AI without letting AI make decisions for the user. Boost suggests one small action only when the user asks for it. Today’s Smile helps the user notice something good in the present, while also allowing “nothing today.”

The public demo starts in English and can be switched to Japanese.

How we built it

Three Notebooks is a Next.js application using the OpenAI Responses API, GPT-5.6, and Structured Outputs. Browser speech recognition creates the transcript; GPT-5.6 performs meaning extraction, structuring, and tentative classification. Data is stored locally in the browser.

Codex accelerated implementation, debugging, regression testing, bilingual support, Git history management, and deployment. The human creator, Grandma, made the core product and interaction decisions: classification by cause rather than date, protecting the value of today, keeping AI suggestions reversible, and requiring a clean zero-start that does not inherit previous input.

Challenges

The hardest problem was not producing a good answer for one example. It was ensuring that a completely different input could start from zero without mixing with previous sessions. We built regression tests around initialization, saving, reloading, incomplete tasks, and duplicate calendar events.

Another challenge was distinguishing a task’s scheduled date from its true origin. For example, doing laundry tomorrow may still belong to the past if accumulated laundry is the reason the task exists.

What we learned

A productivity tool should not only help people do more. It should reveal where their limited attention is going and help them reclaim the ability to live today while still creating a future.

Build Week work

Before Build Week, the project already had the basic three-column interface, manual tasks and calendar items, timeline categories, and local browser storage.

During Build Week, we added mixed voice and text extraction, GPT-5.6 Structured Outputs, editable candidates, cause-based classification guards, contextual White Note conversations, Boost, Today’s Smile, regression tests, bilingual English/Japanese support, and a rate-limited public demo.

What’s next

Future work includes stronger Smile and Boost experiences, optional synchronization and backup, broader accessibility testing, and continued study of how AI can support agency without taking decisions away from the user.

Built With

• next.js
• openai