How to use Cursor to trial Grok on client code
The author found routine edits fast but reported a half-applied migration, failing tests, and JSON debugging that needed tighter steering. They retained Claude alongside the new default.
dev.to
Original prompt
Shared by the source author. See source notes for context.
migrate these three modules and keep generateStaticParams honest
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- agnelnieves
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author found routine edits fast but reported a half-applied migration, failing tests, and JSON debugging that needed tighter steering. They retained Claude alongside the new default.
- Human involvement
- The author stayed in the editor, restored lost constraints, and compared a failed migration with Claude Code.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Set Grok 4.5 as the default model in Cursor for a client week.
- Ran iterative content edits, MCP debugging, and a multi-file migration.
- Logged the raw MCP payload after two client rewrites.
The author made Grok 4.5 the default model in Cursor for a week on a real client repository. The sanitized project was a Next.js App Router publication with content validation, MCP servers for advertising and analytics reads, and an asset pipeline. They kept their existing MCP configuration and used Cursor throughout the trial.
The workload included a feature branch with MDX routes and shared components, a flaky integration returning partial JSON, and a refactor spanning route handlers, a Zod schema, and a TypeScript helper. The author found iterative edits to props, frontmatter, and test stubs fast, and described better handling of partial selections and requests to preserve comments.
The failures were concrete. A multi-file migration updated two files but left another in an intermediate state, with failing tests. Claude Code handled the same request more cleanly the following morning. During JSON debugging, Grok rewrote the client twice before logging the raw payload. In a long thread, the author had to paste the schema refinement rules again. They kept the cheaper default for ordinary editing while retaining Claude for harder work.
Who it's for: Developers evaluating a model against their own repository's editing and integration tasks.
Prompt
Original prompt
Source: dev.to · agnelnieves
Frequently asked questions
What failed in the refactor trial?
Two files were updated while a third remained in an intermediate state, and tests were red. Claude Code completed the same request more cleanly the next morning.
How did the author handle lost constraints?
After a long thread with a large schema dump, they pasted the Zod refinement rules again.
More details
Remaining steps
- Re-pasted schema rules and compared the migration with Claude Code.
About this account
The model edited client code inside Cursor but required steering during debugging and multi-file work.
Platform: dev.to. Attribution: https://dev.to/promptway/i-made-grok-45-my-default-coding-model-for-one-client-week-4gb7. Prompt: Original (quoted from source). Sourced by the Agents In Use pipeline.
Credit: agnelnieves on dev.to
Source published
Added to Agents In Use
Source checked
Tools and topics
Cursor · Grok 4.5 · Next.js · MCP · Zod · Claude Code
Cursor · Model evaluation · MCP debugging
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.