REA + Claude Code: Rebuild One Feature—and Prove It Works

Rebuild one behavior you can describe

The task is search within an app the producer owns. A useful contract says which fields a query can match, how collection filters combine with it, and how results are ordered. That narrows both the implementation and the review. The episode uses REA to gather findings and keeps unknown behavior visible before preparing the Claude request.

Use the awkward cases to define correctness

A single successful query leaves important behavior untested. The episode checks an accent mismatch, words found across different fields, combinations of filters, and stable ordering. Those inputs make the contract concrete: the candidate must return the intended records in the intended order. The deliberately simple control helps show what the fixed cases detect.

Read the result with its method attached

The published experiment reports 24 of 24 cases for the original and the Claude candidate, compared with 16 of 24 for the authored simple control. It also reports five input-integrity and ordering checks. REA was run locally by the producer; Claude received a summary of the findings. Returned code was captured and its syntax spacing restored before execution.

Reuse the method within the same limits

The producer had readable source for this synthetic app, so this is a controlled reconstruction exercise. The control is not presented as a Claude failure, and the result is not a blind benchmark. Use the approach for a feature you own and can observe, then add the security, performance, integration, and language coverage your real application requires.

Jump to a chapter