Propel: A Research Coding Assistant for Claude Code

Propel is a research coding assistant for Claude Code — not an autonomous scientist. Unconstrained LLMs tend to write "the mean of their training data" rather than the specific variant a paper describes. Propel adds structure instead of cleverer prompts: Claude must investigate before it writes code, stop at human-in-the-loop gates to ask design questions, and pass every change through auditors that check it against the paper and hunt for silent bugs. Every research decision stays with the researcher.

The Propel pipeline: seven stages, five human gates, and two questioners.

Key features include four session modes (Researcher, Engineer, Debugger, Trainer), paper-alignment and silent-bug auditors that dispatch after every code change, investigation-first living READMEs, and an experiment registry that carries lessons into future sessions. This website itself was built using Propel with Claude.