The pre-LLM-effort-was-a-filter argument holds up, but I think what effort was really filtering for was why someone built the thing. High effort filtered out “this seemed fun for a weekend” projects. LLMs just surfaced that those were always the majority.

The better filter is: does this project serve a specific audience that genuinely needs it, or is it a demo of what you can do with Claude?

What I look for now:

• Specific problem for a specific group of people (not “general-purpose LLM wrapper”)
• Open core (MIT or something that lets the community carry it if the author walks)
• Revenue model or institutional backing (someone has to keep the lights on)
• Evidence the author understands what they shipped, not just LLM output committed wholesale

We build CircuitForge, self-hosted tools for navigating opaque systems (job markets, government benefits, insurance). The architecture is deterministic-first: eligibility checks, validation, and data pipelines are rule-based and grounded in structured data, so the LLM is drafting from a clean, repeatable foundation rather than hallucinating into a void. That also means we can run smaller, specifically fine-tuned models instead of throwing a frontier model at everything and hoping for the best. Smaller models run on consumer hardware, which cuts hosting cost and shrinks the privacy risk surface significantly. Humans approve before anything acts. Pipeline layer is MIT and lives on Forgejo. There’s a full devops stack, a real business model, and I use these tools every day. We’re also actively collaborating with other devs and always looking for contributors.

The people using these tools actually need them. That’s the commitment signal that doesn’t evaporate when the novelty wears off.