Introduction
And suddenly I’m not overwhelmed by Slack anymore, and I’m less a bottleneck for the teams.
I can focus on what actually matters, while my AI alter ego answers Slack messages and appends and prioritizes my list of fires to put out.
Here’s how I built an AI assistant to manage Slack threads, prioritize work, and reduce context switching.
Problem
It all started when we decided to parallelize many features with my team. A bold move because it meant that I would set the direction and constraints for the product to be built, but the owners of the actual features would be designers, engineers or tech leads.
On paper it was a great idea, but I was still the go-to for the final 10%—the hard decisions. This created a bottleneck. Meanwhile, I was also launching the MVP, talking to customers, handling logistics, legal, customer support… and Slack became a nightmare.
It wasn’t the volume of work—it was the cognitive fragmentation:
- Endless context-switching
- Constant firefighting without meaningful progress
- Forgetting decisions I made just days before
So - how did I delegate to an agent ?
Methodology
Step 1: analyze one by one all the messages I sent in the past weeks, as well as threads in the feature channel and launch channel. For each message I noted: what is the problem, how did I solve it as a human, could a bot solve it ?
Step 2: prioritize the use cases
Step 3: implement use cases
A mindblowing (but retrospectively obvious) learning was: sometimes I get questions from engineers about previously made decisions. First reflex is to think that building an agent with good documentation would solve the issue. But then you realize that if I had written a good documentation in the first place, the engineer would probably not have asked the question.
Cognitive Architecture
I converged quickly on 2 principles:
- The architecture should obviously be multi agent
- The architecture should be modulable - so that I can start by one use case and then unblock other use cases one after the other
The vision was the following structure:
- Supervisor: decides which agent to trigger
- Knowledge agent: fetches context from Notion, Slack history, docs
- Task agents: processes incoming messages into the right actions
This could be implemented in several steps:
- Step 1: only on knowledge agent - no supervisor
- Step 2: a supervisor, the knowledge agent, and a tasks agent
Prompt engineering
I started with basic prompts:
- Step by step instructions for the supervisor
- Simple prompt for the knowledge agent
Evaluation
I tested the system with questions about the product, pricing, status of the ongoing features. The agent got everything right.
The most difficult part will be to avoid answering to all messages that will be sent in the slack channels - to be tested starting Monday
UX
The idea for the UX is very simple: what if I could duplicate myself ? So there will just be another PM in the Slack channel called Daasarp (did you get it ?) with a profile picture of myself upside down.
The biggest challenge will be of course to add supervision - in an ideal world, the agent would observe my answers, learn from them, try to answer some threads, and get feedback. An easy UX would be to add emojis to answers (mine and the agents) to educate about good and bad answers.
Conclusion
This is only the first step of being an augmented PM with AI. In theory, all the context I need to do my work is limited to Notion pages, Slack and meeting recordings - that could be consumed by AI. The potential for replacing whole parts of my work is huge.