claire vo
← all AI workflows

Prototype an agent-populated client workspace with Codex

A central project folder connects a conversation, decision card, and completed work object on a small planning table.
prototype-an-agent-populated-client-workspace.png
A central project folder connects a conversation, decision card, and completed work object on a small planning table.
Codex + Connected Granola, Gmail, Google Drive and CRM evidence + Browser prototype review

I asked Codex to prototype a CXO.dev workspace from the engagement context available in my connected tools. I wanted to see the work with each client: the conversations, initiatives, decisions, and deliverables that belonged together.

The first versions felt like an internal administration dashboard, with disconnected records and too much commentary. I pushed the design toward the initiative itself. Opening an initiative should show the conversation that shaped it, the decision we made, and the next piece of work.

I also made the intended input method explicit: agents would populate the records. I was not asking for another set of forms to maintain. Codex removed the manual entry controls and connected the prototype around an agent-assembled snapshot, with sources available behind the records.

It was still a prototype; continuous syncing was not connected. That boundary mattered, but the prototype let me review the structure before building the ongoing system. The most useful feedback was about relationships and how the work arrived, rather than asking for a more elaborate dashboard.

Promptprototype-an-agent-populated-client-workspace.md
# Prototype an agent-populated work view

Use only the sources I authorize. Identify the initiatives, conversations, decisions, actions, and deliverables supported by those sources. Keep source links, dates, and uncertainties with each record. Do not fill gaps with invented activity.

Organize the view around what we are doing. Each initiative should connect its relevant conversations, decisions, next actions, and outputs. Avoid isolated lists that require me to reconstruct those relationships.

Assume agents will populate this workspace. Build a readable review surface rather than manual data-entry forms. Make conflicting or missing evidence visible without surrounding every record with implementation commentary.

Use private data only in a private prototype. Prepare a separate synthetic example if a public demonstration is needed.

Let me review a small connected example before expanding the interface. State which parts are an assembled snapshot and which, if any, have continuous syncing. Do not imply an ongoing integration from a successful prototype.