tight ingot
Hi everyone,
I’ve been working on an open-source project called ProofHound:
https://github.com/proofhound/proofhound
It is a platform for optimizing prompts against real classification datasets (generation and agents will be supported soon). The goal is to make prompt improvement more systematic instead of relying on manually tweaking prompts one by one.
At the current stage, ProofHound focuses on classification tasks and supports a workflow around:
Running prompts against labeled datasets
Comparing prompt versions with evaluation results
Automatically optimizing prompts based on failure cases
Managing prompt versions across the lifecycle
Moving from experimentation toward release and governance
The broader direction is to build a full prompt lifecycle platform: from debugging and optimization, to version management, evaluation, release, monitoring, and future support for more task types beyond classification.
I’m building this because I think many teams still manage prompts in a very manual way, especially when prompts are tied to production business logic. A dataset-driven workflow can make prompt iteration more measurable, repeatable, and easier for non-engineers or prompt owners to participate in.
If this sounds useful, I’d appreciate any feedback, GitHub stars, or suggestions.
You can also join the Discord if you’re interested in following updates or discussing use cases.
GitHub: https://github.com/proofhound/proofhound
carmine forge