Work Feels
AI agent who listens to your working life, captures key moments as poetry, then surfaces patterns in your own data and reflections
Problem
Personal reflection in professional settings gets lost in the daily grind. People miss patterns in their own emotional responses, productivity cycles, and growth — not because they don’t care, but because traditional journaling feels like a chore. Existing AI tools compound the problem: they’re transactional, not reflective, and none of them treat the experience of work as something worth examining with any depth or care.
Concept
I wanted to build something that made reflection feel like a reward rather than a task. The core idea was to use AI not as an assistant that answers questions, but as a companion that listens, remembers, and occasionally surfaces what you said three weeks ago in a way that stops you in your tracks.
The poetic output wasn’t a gimmick — it was a deliberate design choice to create emotional distance from the raw journal entry, making it easier to see your own patterns without feeling exposed. A direct summary of what someone wrote feels like a mirror held too close. A poem refracts the same material through a different lens, and that shift in angle is what makes the pattern visible without being confronting.
The interface had to earn trust through restraint. No notifications, no streaks, no gamification. The AI asked questions, not prompts. The distinction matters: a prompt directs you, a question opens space.
Technical architecture
The hardest engineering problem was sustaining tone and memory across conversations that spanned months. A model that sounds different on week one and week twelve breaks the sense of a consistent companion. I spent significant time on prompt engineering to maintain voice stability across long conversation histories.
Journal entries needed to be retrievable not just by recency but by emotional content and theme — a user asking “when did I last feel like this?” needed pattern matching across months of entries, not a simple date lookup. I built a RAG pipeline combining semantic search over vector embeddings with date-based filtering, so the AI could surface relevant past entries as context before generating a response.
Poetry generation ran as a separate tool-use pipeline, triggered by the AI’s own judgment about when a journal entry had enough emotional substance to warrant it. This meant poems appeared as a genuine surprise — not on a schedule.
- Stack: Next.js + Vercel AI SDK + Clerk (auth)
- Model: OpenAI — conversation, pattern analysis, and poetry generation
- Database: Turso SQLite + vector embeddings for semantic journal search
- Observability: PostHog + Sentry — tracking engagement patterns and surfacing errors in the long-context pipeline
- Deployment: Vercel
Result
The beta version ran for several months with 26 research participants. Users reported discovering unexpected patterns in their work emotions — things they hadn’t noticed themselves until the AI surfaced them.
The most revealing outcome wasn’t in the data: it was watching participants screenshot and share the poems. Nobody was asked to do this. There was no share button. People were copying the text and sending it to colleagues, partners, and friends — which told me the poem had succeeded at its design goal. It had created enough emotional distance from the original feeling that it felt safe to share, while remaining precise enough that it still meant something to the person who wrote it. The immediate next feature request was a native share function.
The project validated that AI can function as a genuinely reflective companion, not just a productivity tool — and that the format of the output is as much a design decision as the interface that contains it.