A course correction is a follow-up instruction that redirects a coding agent after its output reveals an unwanted assumption. It can bring a working result closer to the intended behavior, but it leaves the agent dependent on continued human inspection and steering. section 2.3.1 section 2.3.3
Correcting a URL shortener
An initial request asks for a dependency-free Node application that turns long URLs into short URLs and redirects visitors. After testing the redirect, inspect choices the request left open: the agent might use six-letter codes, generate a different code for the same URL each time, and store mappings in memory. A concrete follow-up is: “Slugs should be 7 characters, mixed case alphanumeric. The same long URL should always return the same slug.” Keep the active session so the follow-up retains its context. section 2.3.1
Verify the corrected behavior before requesting more work, such as a URL listing, a swappable storage layer, or slug-generator tests. Repeated correction is a poor substitute for stating stable requirements early: each omitted decision can create another inspection-and-redirection turn, while the growing conversation must be sent back to the model. section 2.3.1 section 2.3.3
