How I built an AI agent that keeps me connected to product managers, without another meeting or another status report.
- My role: Designer and builder
- Built with: Microsoft Copilot Studio
- Reads: Teams, Outlook, Azure DevOps, and GitHub
- Scope: My four Azure AI content areas: Foundry Local, Translator, Language, and Document Intelligence
The problem
Recurring syncs with product managers stopped working as my main way to stay close to product direction. Meetings got canceled or rescheduled, time zones added friction, and many PMs preferred to work asynchronously. Meanwhile, the context I needed (decisions, risks, roadmap changes, customer pain points) was scattered across chats, email, work items, and pull requests.
More meetings wouldn’t fix that. The meetings were never the real problem; the awareness model was. So my question wasn’t “How do I build an agent?” It was: Can AI help me connect those signals and stay meaningfully engaged with my partners?
Five questions
The breakthrough was realizing that almost every valuable conversation with a PM comes down to five questions:
- What changed?
- What do I need to know?
- Is anything blocked?
- Do we need something from each other?
- Where are customers struggling?
Once those questions appeared, the project stopped being “How do I monitor everything?” and became “How do I consistently answer these five questions?” Each of my content areas got a standing Content Status space in Teams, built around them.
What spot does
spot is a lookout, not a dashboard. It sweeps my connected sources and turns scattered signals into actionable context: what changed in an area, what may need PM attention, which follow-ups are still open, and where customer or content friction is starting to show.
- It acts. Observations become outputs: a drafted message, a flagged risk, a recorded decision.
- It works in both directions. It finds what I owe people and what people owe me.
- It looks for gaps. Often the most valuable thing it surfaces is the question nobody has asked yet.
- It never sends. spot drafts; I press send. Always.
Design decisions
- Natural language always works. Commands like /pulse, /three, and /snapshot are shortcuts, never requirements. An agent that needs the right incantation is a tool you have to operate; one that answers a plain sentence is a colleague.
- Memory outranks the instruction page. Each memory file holds one fact, so every correction lands somewhere findable instead of diluting a single long document.
- Live sources first, memory second, never fabricate. Every report closes with what it actually swept, which makes silent connector failures visible.
- Silence is a valid output. An honest, near-empty report beats a padded one.
- Teach it what “done” means. A merged PR isn’t published, and a 👍 reaction is a real reply. Each “no, that’s not what that means” becomes a rule.
The feature I’m proudest of: /postscript
A postscript is the thing you add after you thought you were finished. Early on, spot noticed that a documentation fix had shipped, but the person who reported the problem had never heard back. It didn’t find a task; it found a missing human connection.
/postscript looks for commitments that are closed in my head but open in fact. It orders them by how long the other person has been waiting, not by importance, and drafts the message that closes each loop. Only a sent message closes a loop: not a merge, and not a memory.
What I learned
- Lead with the problem, not the tool. People connect with the awareness gap long before they care about the agent.
- Start with fewer commands than you think. The first version had 16, and most went unused. A command on the page is a promise.
- Setup takes minutes; getting it good takes about two weeks. The agent improves by being corrected out loud, every day.
Sharing it
I demoed spot to my team and turned it into a portable build specification and a copy-and-paste kit, so other content developers can build their own awareness agent in about fifteen minutes.
