//context
Teams use AI heavily but lose value because context lives everywhere except inside the AI conversation. Every chat starts from zero: re-explaining the product, the customer, what has already been decided, and how the team works. Contexx sits between Notion and Cursor: a workspace for mixed teams, not just engineers.
The origin was a real, lived problem. At Pearson, we started using a markdown and GitHub context repository for research work. The idea worked, but the GitHub setup, lack of multiplayer, and local-not-cloud workflow made it hard for non-technical teammates to adopt. Contexx was my answer: deliver the same context-grounded AI benefit without that friction, cloud-based, multiplayer, and approachable for non-engineers.
//build facts
Commits
107
Timeline
Feb-Apr 2026
Build model
Solo, built in Cursor
Research
15+ interviews
//key design decisions
The design challenge was agent UX at real-time scale: make context visible, reusable, and controllable without magic. The substance was not a chat UI; it was the system of decisions around Ask vs Agent modes, streaming responses, and token budgets that makes AI behavior understandable for a team.
Three-panel workspace
A document tree, editor, and AI chat sit side by side so users can read source material, edit documents, and talk to AI without switching modes or tools.
Proposal-gated AI editing
Agent Mode can propose new documents or edits, but every meaningful change requires user acceptance through explicit review states. Editing the diffs directly was the next iteration I had planned before pausing the project.
Context first, prompting second
The product assumes better AI output should come from well-maintained context, not from users becoming better prompt engineers. Users can select text in a document, attach documents with @, and invoke skills with /.
Document vs Skill
Contexx separates passive knowledge from active instructions: documents define what the AI should know, while skills define how it should behave.
Visible context, not magic
Users stay informed about the context, skills, and thinking the agent uses while it runs.
User Skills vs Project Skills
My Skills live in a separate section and work across projects. Project Skills live inside a project folder, but can be added to a user's Skills when they become reusable beyond one workspace.
//process
There were no Figma files. I built Contexx entirely in Cursor and Claude Code, solo, and let the product evolve through use. I started from a three-panel IDE layout: browse files on the left, read them in the center, prompt the agent on the right.
To move fast and prove the concept, I leaned on open-source libraries instead of rebuilding everything from scratch, then started loading documents, editing them, and prompting the AI against them. From there it grew feature by feature, beginning with a simple Ask mode and a token budget to keep costs sane.
The real engine was dogfooding: I used it daily, watched how the AI behaved against the rules and context I set, and shaped the experience from there. After a few weeks it had become a tool I reached for constantly, and a build process that taught me more than any spec could.
//my role
I owned the product thesis, information architecture, interaction model, and visual design, and built it end to end in Cursor and Claude Code. I did not hand a spec to engineers; I drove the whole thing solo with AI tooling, from the first idea to a working alpha.
//impact
I dogfooded Contexx daily and improved it in the same loop, building, using, and refining in one cycle. It was also validated through the Pearson research-repository workflow, where context-grounded AI worked well but a custom corporate tool raised the expected risks around sensitive data, governance, and adoption.
//challenges
The hardest constraint was token cost. I cut it by more than 10x with a retrieval (RAG) layer and by splitting models by job: a cheaper model for querying and retrieval, a stronger one for answering. I also experimented with an Auto mode that switched models based on the type of request, but never got it reliable enough to keep.
In the end I stopped building Contexx; as a solo project it grew hard to maintain and extend. We trialed it at Pearson, but the upkeep plus the risk of holding personal data there made it not worth it, so we defaulted to a Cursor and GitHub setup for team work.
I'm glad I was early, though, the wave of startups now building in exactly this space says the bet was right.