Complex code modification via chain of thought prompting

A computer-implemented method comprising:

• receiving, at modification service implemented in a cloud provider network, a request to modify a user-provided code segment according to a modification goal, the request including or identifying the user-provided code segment;
• obtaining, via use of a database, a set of modification steps, the obtaining comprising:
• computing similarity values, wherein each of the similarity values is indicative of a similarity between a representation of the code segment and representations of other code segments;
• identifying a largest computed similarity value from the computed similarity values, wherein the largest computed similarity value is indicative of a similarity between the representation of the code segment and a representation of a first other code segment;
• and
• obtaining the set of modification steps from a record of the database corresponding to the first other code segment;
• generating a modified code segment based at least in part on providing the code segment, one or more of the set of modification steps, and an example modification to a machine learning (ML) model, wherein the example modification includes data associated with the first other code segment;
• and
• transmitting the modified code segment for presentation to the user.