The $280-per-hour contractors teaching Claude to code better reveal the messy human scaffolding holding up the agent economy.
The Summary
- Anthropic is using ~1,000 human software engineers via Snorkel AI to fine-tune Claude Code, paying contractors $280/hour to A/B test code outputs and train the model to write cleaner, more maintainable code.
- Project Marlin (Snorkel's internal name) focuses on teaching Claude Code to mimic professional developer patterns, not just produce working code.
- This reveals the labor structure behind AI coding tools: human experts teaching machines to replace... human experts.
The Signal
Anthropic's Claude Code isn't getting better by magic. It's getting better because roughly 1,000 software engineers are spending hours teaching it what good code looks like. The project, run through data labeling firm Snorkel AI, pays contractors $280 per task to create prompts and compare outputs from two different models, choosing which version better matches professional developer standards.
The work is straightforward: A/B test code snippets, pick the winner, explain why. Each task takes about an hour, though some submissions bounce back for revision. The goal isn't just functional code. According to project guidelines reviewed by Business Insider, contractors are training Claude Code to write simplified, maintainable code with the level of detail a senior engineer would expect.
"The project aimed to ensure the model could achieve the level of detail expected in the prompt, essentially training Claude Code to write simplified, easier-to-maintain code."
This is the economic loop powering the agent economy:
- Expert humans teach AI to do expert work
- AI gets good enough to threaten expert human jobs
- Expert humans need new work, maybe teaching the next model iteration
The $280 hourly rate matters. That's senior developer money, not mechanical turk wages. Anthropic isn't buying cheap data labeling. They're buying professional judgment about what separates good code from working code. The contractors need software engineering backgrounds to even qualify. You can't teach taste without having it first.
What's striking is the outsourcing structure. Anthropic doesn't hire these engineers directly. They contract Snorkel AI, which hires the contractors. This creates distance between the AI company and the human labor improving their models. It also reveals a growing industry of specialized data firms that know how to source, vet, and manage expert contractors at scale.
The contractors interviewed didn't know which model versions they were evaluating. That opacity is standard practice in AI training to prevent bias, but it also means the people improving Claude Code can't see their own impact in real time. They're training a system that might make their skills less valuable, one blind comparison at a time.
The Implication
If you're a software engineer, understand the game board. AI coding tools aren't emerging from pure algorithmic progress. They're being hand-tuned by people who do what you do, using the same judgment you use. The better Claude Code gets, the more valuable that judgment becomes for training, even as it becomes less valuable for writing production code.
For companies building agents: quality training data comes from quality humans. The race to better AI isn't won with bigger models. It's won with better feedback loops from people who actually know the domain. Budget accordingly.