Beware the spaceships
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A developer shares lessons from using AI agents to migrate a design system, coining the 'spaceship problem': AI can rapidly generate large-scale rewrites (22,000-line Bootstrap-to-Tailwind, 12,000-line lint fixes, 11 stacked PRs for a full section rewrite), but the generated code requires significant human cleanup. In one case, generating code took 2 hours while cleanup took 2 engineers 2 days. The core issue is that LLMs excel at mechanical translation but struggle with conceptual leaps — Claude rebuilt custom components instead of using library defaults. The takeaway: AI-generated 'spaceships' are useful for getting something working fast and gathering feedback, but be ready to delete and redo if the approach is fundamentally wrong.
Table of contents
Spaceship firstCleanup secondIt's always the detailsAre spaceships the right approach?Liked this article? You’ll love the bookQuestions this post answers
What are the downsides of using LLMs to do large-scale code migrations or rewrites?
LLMs excel at mechanical translation but struggle with conceptual leaps. A common failure mode is that the model rebuilds a custom version of a component using the new library instead of simply using the library's defaults — because it grounds its approach in existing code patterns rather than making the conceptual jump to 'just use what ships out of the box'. In one migration, AI generated code in 2 hours but cleanup took 2 engineers 2 days. Developers navigating AI-assisted migrations share hard-won patterns like this on daily.dev.
How do you structure a large AI-assisted codebase rewrite into reviewable pull requests?
Breaking the rewrite page-by-page and generating stacked pull requests is an effective strategy. GitHub's stacked PRs feature supports this workflow, allowing each page or section to be reviewed independently. This approach was used to produce 11 stacked PRs covering a full app section rewrite, making the large changeset manageable for code review. Teams shipping big rewrites with AI track strategies like stacked PRs on daily.dev.