Anthropic has deployed a large-scale human feedback operation to sharpen Claude Code, its AI coding tool that has recently disrupted the vibe-coding industry. The effort, code-named "Marlin" inside Snorkel AI, leverages roughly 1,000 software engineers to fine-tune the model's outputs. Those contractors are paid up to $280 per task to create prompts and evaluate responses.
AI firms routinely outsource data work to third parties like Snorkel, which hires contractors globally to teach models specialist subjects. Project Marlin specifically aims to make Claude Code mimic the skills of a professional developer. The initiative reflects the growing reliance on human-in-the-loop systems to bridge gaps in AI performance.
Contractors working on the Anthropic project reported earning $280 per completed task. The work involves creating prompts and assessing Claude Code's outputs, a process that feeds directly into the model's fine-tuning pipeline. Snorkel AI is managing the contractor pool and training materials for the project.
By investing in targeted human feedback, Anthropic is betting that quality data from experienced engineers will give Claude Code a competitive edge. The approach could accelerate adoption among professional developers who demand reliable, production-ready code generation. It also raises questions about scalability and cost as demand for specialized AI training grows.
Critics argue that relying on high-wage contractors makes such fine-tuning economically unsustainable at scale. If Anthropic cannot automate quality assurance, the model's improvements may plateau once the initial training rounds conclude.