recently, business insider reported that anthropic is advancing a specialized optimization initiative codenamed “marlin,” aimed at significantly enhancing the practical development‑readiness of its coding assistant, claude code. the project is being fully executed by the data‑annotation service provider snorkel ai, bringing together roughly a thousand software developers with real‑world engineering experience. through an intensive, scenario‑driven feedback loop, the effort seeks to align the model’s outputs more closely with industrial‑grade coding practices.
unlike general‑purpose data annotation, the marlin project focuses on fine-tuning the model within highly realistic engineering contexts: outsourced engineers are tasked with crafting structured prompts based on actual development tasks, reviewing the quality of generated code, and conducting blind comparisons between the outputs of two parallel models. they must not only determine which piece of code best matches the intended prompt but also assess its readability, maintainability, and level of detail—ultimately aiming to “enable claude code to write cleaner, more robust code that aligns more closely with human engineers’ intuitive approaches.”
participants revealed that each task carries a compensation of $280 and typically takes about 60 minutes to complete; however, due to stringent review standards, some submissions require multiple rounds of feedback and revisions from snorkel. notably, all evaluations are conducted under conditions where model version information is completely obscured, ensuring objective and impartial feedback. at present, the project remains in an ongoing iterative phase.
