The product craft is changing fast. Fast enough that parts of the job didn’t exist a couple of years ago. PMs now prototype in code, wire up agents, and write their own evals. So I gave myself a runway to do that. Everything below I designed, built, and shipped solo, as both the PM and the engineer.
Here’s the tour.
distribution-agent

Here’s the idea I couldn’t stop thinking about: building has never been easier, so the hard part has quietly moved to getting seen. Every solo founder I know can ship a product in a weekend and then has no clue how to find their first hundred users.
So I built them a teammate. distribution-agent reads the rooms where their customers already hang out, surfaces the threads worth joining, and drafts a reply in their voice. Then it stops. The founder reads it, tweaks it, and hits post. The agent never publishes on its own (violates T&S conditions).
Under the hood it’s a Next.js app on Supabase with durable background jobs and a real eval harness, including a prompt-injection red-team so a hostile thread can’t hijack the draft. I wanted it to feel like a sharp junior teammate who does the legwork and always checks with you before speaking.
pm-brain

The agents I build kept handing me confident, generic product advice, the kind you’d get from a listicle. I wanted a thinking partner that had actually read the good stuff.
pm-brain is that partner: a live retrieval system loaded with a curated library of product-strategy frameworks. When I’m poking at a new idea, say, a travel app, I ask it where the moats might be and where the whole thing could fall apart, and it reasons back with real frameworks behind it. My other agents hit the same endpoint when they need to think strategically. It’s deployed and running – a FastAPI service with hybrid search and reranking, exposed to Claude Code as an MCP server.
hermes agent flow

Running four workstreams at once as a team of one gets messy. hermes is the chief of staff I gave myself.
It’s a small crew of agents: one for business, one for research, one for building, with an orchestrator above them that sets the week’s goals, tracks KPIs, and hands work between them. They ping me when something needs a decision, and I approve it over email. The whole thing runs on headless Claude Code with a local model as backup, so it keeps humming even offline.
The fun part was designing the etiquette: when an agent should just act, when it should ask, and how the orchestrator keeps everyone rowing in the same direction.
pm-gates

After a few builds I kept hitting the same failure: the code was clean, the tests were green, and the product still solved the wrong problem. The missing layer was product judgment.
pm-gates is that layer, made executable. Eleven connected gates walk a build from “is this even worth doing” through “how will we know it worked” to “is it safe to ship”. And it stops for a human at the two moves you can’t take back: pushing to production and migrating a database. I published it as an open-source Claude Code plugin, and it’s the loop every project on this page ran through.
If you’re building something ambitious and want a product partner who’ll get into the weeds with you, I’d love to talk.
Email: deepika.gprakash@gmail.com · LinkedIn: /in/deepika-gnanaprakash · GitHub: /deepikabg