Everyone seems to agree that company context is an interesting problem to solve for enterprises, and I'd like to solve the product layer here at ChatPRD.
Unlike code which is largely deterministic and engineering work (PRs, tickets) which are mostly logical, Product data is poorly structured, communicated randomly, and stuck in the heads of product managers who largely disagree with each other.
Building a source of truth on the chaotic mess that is product management is perhaps a Sysyphean task, but that's why we have /goal.
What are the primitives of product management
So at ChatPRD I've been playing with the idea of a Product Graph, and what are the fundamental primitives of product management and can you infer them with relative accuracy from the various signals (external and internal) available via API / MCP / whatever else an agent can brute force against.
I think it probably looks something like this:
You would think that throwing a 1M context window mega model at that bottom layer (data signals and systems) and just saying “yo what is going on here” would get you good results, but guess what? It does not!
Semi-auto research, or 15 loops to nowhere
I figured what better place to start that my own ChatPRD which is scattered across github, linear, intercom, chatprd (ofc), granola, and our data warehouse.
Ingest of this data is pretty easy with the right apis, but making any damn sense of what's going on, even for a relatively small business and team of very few!
While it was fairly easy to spin up a run-of-the-mill insights engine (what people are complaining about), but reconciling that data with a) are we working on it b) why and c) is it working is trickier!
So I threw all the agents and dynamic workflows and /goals at this problem, trying to figure out what combination of architecture decisions would create a high quality pipeline of raw data -> useful graph and aggregation--starting with ChatPRD specific flow and then generalizing to work with any prouct.
I tried lots of things. Here are just a few! Lots didn't work well!
Company-specific knowledge compiler