After I generalized my tool's input to read standard trace files, the obvious next step screamed at me: keep going. Build adapters for n8n, for Dify, for people who wrote their agents in plain Python. Cover everything. More coverage, more reach — right?
I stopped, because that instinct is exactly the one that killed my first project: building features nobody had asked for, polished to perfection, used by no one.
The split that made the decision for me: there's a standard ecosystem and a non-standard one. LangGraph, CrewAi, AutoGen, LlamaIndex all emit OpenInference traces — so one entrypoint opened all of them at once. Building that was demand-driven: I'd hit the real friction of "I can't get your trace in" with actual people. That's a wall the market put in front of me, so removing it is justified.
The non-standard side — n8n, Dify, hand-rolled Python — uses bespoke formats, and each needs its own adapter. Here's the trap: I could build all of them now. It would feel productive. Tests would go green. And not one of them is blocking a real user today, because I don't have a real user on those frameworks asking. Building them up front isn't coverage — it's speculation with a compiler.
So the rule I'm holding: build the door when someone is actually standing at it. The standard ecosystem had people at the door. The non-standard adapters get built the day a real user shows up with that format and that pain — not a day before. Their trace will also tell me exactly what their format looks like, which beats guessing.
"What's the disciplined thing to build next?" often has the answer "nothing yet." The hard part of solo building isn't writing code. It's not writing the code that feels productive but isn't wanted.
Code: github.com/JEONSEWON/Clew-by-Custos
Top comments (0)