Anthropic Engineer: We Stopped Writing Prompts Internally Long Ago

An Anthropic engineer spent half an hour on stage showing how they create self-prompting loops internally. She said: Internally, we barely write prompts anymore; we write loops.

This isn't idle conceptual talk. She directly showed how the Claude team actually does it:

• Memory is no longer a single CLAUDE.md; it becomes a memory/ directory that the agent reads and writes itself
• Memory rots when kept in the conversation: attention gets diluted, patterns get buried, and old notes are still confidently misapplied
• During the day, each agent writes its progress into the team memory
• At night, a round of dreaming runs: checking, organizing, and filling in gaps

The agent that starts work the next day is automatically a little smarter than it was yesterday. The data results are clear: first-pass errors dropped by 97%, and acceptance was 30% faster. People go back to looking at the product instead of constantly staring at prompts.

This approach isn't confined to the talk. Hanako published a detailed tutorial that turns Claude Code into an AI quantitative research team: a lead handles planning, a scout checks the data, a sweeper runs parameter grids, a skeptic is dedicated to attacking the results, and a librarian searches the literature. The team shares a research.md as memory, and every experiment must record a verdict: DATA VALIDATED, TRADE-OFF, or NO EDGE. Only results that the skeptic cannot break count.

The whole system runs on QuantGPT, connecting to Claude Code through MCP to provide historical data, cloud backtesting, and trading interfaces. Hanako emphasized one principle: the person who comes up with an idea should not score their own work. A single chat window can't do this; it is simultaneously player, referee, and scorekeeper.

Prompts are one-off; loops can run continuously. The difference isn't whether you know how to ask questions—it's whether your agent, after a night's sleep, still remembers what it learned yesterday.