I built Hanji to
review and direct
agent work.
AI agents were writing documents and carrying out more of my tasks. I needed clearer ways to read their output, judge it and decide what happened next.
Write made agent documentsreadable and reviewable.
Agent-created documents were easy for agents to use. I needed people to read, question and revise them together. Write became a shared surface for the words, the discussion and the decision to accept a change.
Write · Private test · Request access
People and agents could change the same document. How could I keep my draft and still judge an agent’s proposal?
People decide
which changes belong.
I kept the proposed edit beside its author and controls to reject or accept it. Human discussion stays attached to the passage being judged.
When an agent changed a document I was editing, preserving my draft became the priority. Independent edits can merge; overlapping changes need a choice.
Work let agents executewhile I judged the plan.
Readable documents were only the start. I needed to see how the work connected. In Work, agents create issues, organize views and carry out tasks. I inspect the whole picture and direct them at workspace or issue level.
Work · Early build · Request access
Instructions in a separate conversation lost the context of the task. Where should the conversation live?
Commands moved
into issue comments.
I tried terminals, then a separate in-app conversation. Bringing the AI assistants Claude and GPT into the task’s comment thread kept the instruction, response and next decision beside the work.
Organizing a task should not silently trigger an agent. I replaced label-driven requests with explicit commands.
Think connects sketchesto documents and tasks.
My early thinking still lived in Figma and FigJam. In my setup, FigJam’s agent connection did not give me the control I needed, and moving ideas into product management meant plugins and workarounds. I started Think to connect that stage to Write and Work.
Think is a local prototype.
Carry the idea into execution
without losing its source.
These connections are implemented locally and still being tested together.
Daily use shapeda calmer, faster interface.
I use Write and Work at Buzzvil every day. That made small frictions hard to ignore: unreadable color, a lost cursor, an unnecessary wait. The visual system and the interaction details grew together.
Hanji
Write builds its light and dark color schemes from three starting colors. Work uses carefully selected purple palettes. I tested both in the interface: text, transparent highlights, focus and action.
Daily use made small interruptions add up. Moving between reading, editing and directing needed to stay quick.
Repeated actions
need fewer interruptions.
I kept reading-to-editing inside the same page and prepared the editor on hover or focus. Opening another Work issue preserves the board underneath.
Shortcuts grew from the actions I repeated. They stay out of the way when someone is typing.
- Find an issue
- ⌘ / Ctrl K
- Reach its comment
- c
- Send an instruction
- ⌘ / Ctrl Enter
What I learned:judgment needs context.
A document, an issue and a sketch each make something easier to see. The important design work was keeping the reasoning intact as it moved between them and keeping people oriented while agents changed the work.
Daily use gives me a close feedback loop. The next step is to extend it to roughly five to ten people, and learn which decisions hold up outside my own workflow.
Write is being tested privately with a small group. Work is an early build, and Think remains a local prototype. Open source is a direction; there has been no public source release.