//context
At Pearson, our research workflow involved repeatedly explaining the project, customers, and past decisions to AI. We started keeping that context in Markdown files on GitHub, but the local setup was difficult for non-technical teammates to use.
I designed and built Contexx as a shared workspace where a team could keep documents, edit them, and use them in AI conversations. The working alpha combined a document editor with chat and reviewable AI edits.
//key design decisions
I designed Contexx around two questions: what is the AI reading, and what can it change? Users can see which documents inform an answer. Ask mode answers questions; Agent mode proposes document edits that users review and approve before anything changes.
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.
Review AI edits before applying them
Agent mode proposes documents and edits for the user to accept or reject. I planned to let users edit a proposal directly, but paused the project before building that step.
Bring documents into the conversation
Users can select a passage, attach documents with @, and invoke reusable instructions with /. This keeps the source material close to the question.
Document vs Skill
Contexx separates passive knowledge from active instructions: documents define what the AI should know, while skills define how it should behave.
See what the AI is reading
The chat shows which documents and skills the agent is using while it works.
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
I built the alpha in Next.js with Supabase, Tiptap, and the Vercel AI SDK, using Cursor and Claude Code. Open-source libraries let me test the document and chat workflow early.
I used it daily with my own documents, checking whether the AI followed the attached context and instructions. That gave me specific problems to work on, from selecting source material to reviewing an edit without losing my place in a document.
//team trial
We trialed Contexx at Pearson after using a GitHub repository to give AI access to research context. The workflow was useful, but maintaining a separate app and handling sensitive data introduced more responsibility than the team could take on. We returned to Cursor and GitHub.
//challenges
Token costs became a constraint. I added retrieval and used a cheaper model to find relevant context, reserving a stronger model for answers. I also tried automatic model selection, but removed it when the routing proved unreliable.
I paused the project when maintenance and data-governance requirements became too much to manage alone. A further version would need a clear owner for those responsibilities before a broader team rollout.