__
/\ \
__ _ __ __ \_\ \ __
/'_ `\/\`'__\/'__`\ /'_` \ /'__`\
/\ \L\ \ \ \//\ \L\.\_/\ \L\ \/\ __/
\ \____ \ \_\\ \__/.\_\ \___,_\ \____\
\/___L\ \/_/ \/__/\/_/\/__,_ /\/____/
/\____/
\_/__/# personal RLHF. without the H being a stranger.
Train your agent by labeling its work.
Label your AI's outputs. Compile the labels into how it works.
Your agent writes markdown with prose, code, and mermaid. You label what's off — on the exact line, span, or block — or answer multiple-choice pages. Your agent reads the labels through MCP and the next one is closer.
ship → label → learn → improve
- ›code stylelabel inline samples of AI-generated code.
compile into CLAUDE.md / cursor rules. - ›writing voicelabel what doesn't sound like you in the AI's draft.
compile into a voice guide. - ›second opinionlabel what's missing in the AI's research or analysis.
compile into research guidelines.
› examples
- › private by default · ephemeral by default
- › zero passwords · remote MCP at
mcp.grade.dev

